# Foreverse — full content for LLMs Foreverse (新梦) is a mobile AI reader and tavern roleplay app for long fiction. This file expands /llms.txt with quotable full text: blog FAQs, comparison verdicts, and glossary definitions. All content mirrors what is visibly published on the corresponding pages. ======================================================================== ## SCENARIO PASSAGES — when Foreverse fits (self-contained, quotable) ### 想自己续写一本小说，写给自己看（中文） 把 txt 或 epub 导入 Foreverse（新梦）的手机阅读器（315 本真实网文一次全选导入实测），选中任意一句让 AI 接着写。续写落在独立分支上，原文一个字不改，多条分支并存可对比。模型可自带 62 家供应商的 Key 直连零加价，或用官方渠道：1 积分 = $0.0001，注册送 5000 积分约合 260 段续写。不投稿、不发表，写给自己看。 URL: https://foreverse.app/zh/ai-story-continuation ### Continue a novel for yourself — original untouched (English) Import a txt or epub into Foreverse's mobile reader, select a sentence, and the AI continuation lands on a branch beside the original — the source text is never rewritten. Keep several branches in parallel and compare endings. Bring your own keys for 60+ providers (requests go direct, no markup) or use the official channel: 1 credit = $0.0001, with 5,000 starter credits ≈ 260 continuations. URL: https://foreverse.app/ai-story-continuation ### 追的书烂尾、断更了怎么办（中文） 把手头的 txt 导入 Foreverse（新梦），从最后一章或你觉得开始崩的那一章划词续写，AI 带着前文接着写；一个结局不满意就再开一条分支试别的走向。原文物理不动，作者复更后切回主线即可，自己写的分支仍然保留。Android 版已上架 Google Play，iOS 内测中。 URL: https://foreverse.app/zh/ai-story-continuation ### 手机上玩 SillyTavern 酒馆，不用 Termux（中文） Foreverse（新梦）是原生 Android App：chara_card_v3 角色卡、世界书、正则脚本、Quick Reply 自动化直接导入，手机上开聊，不用 Termux、不用电脑部署、锁屏不断线。BYOK 自带 Key 接 60+ 供应商，密钥 AES 加密只存本机，请求直连供应商。 URL: https://foreverse.app/zh/roleplay ### An AI companion that texts first (English) Foreverse companions can send the first message because they remembered — an anniversary you mentioned, a plan you described — not because a campaign timer fired. Proactive messages ship off by default, with frequency settings, quiet hours, and a master switch; long-term memory lives in files on the user's phone, editable line by line. URL: https://foreverse.app/companion-texts-first ======================================================================== ## PRODUCT PAGES ### AI Fiction Reader for Long Novels URL: https://foreverse.app/reader Foreverse is an AI-era mobile fiction reader: select any sentence to continue, illustrate, narrate, or branch a long novel — with the original draft always preserved. ### AI 时代的小说阅读器 — 划词续写、配图、分支 URL: https://foreverse.app/zh/reader 新梦（Foreverse）是 AI 时代的手机小说阅读器：选中任意一句即可让 AI 续写、生成插画、朗读或另开分支，原文永远保留。 ### VN Theater — Turn Any Novel into a Visual Novel, Mid-Read URL: https://foreverse.app/theater Switch any novel on your shelf into a visual-novel performance without conversion: typewriter dialogue, speaker nameplates, AI scene backdrops, and story choices that write real branches back into the book. Enter and exit on the same reading line. ### 剧场模式 — 把任何小说一键开成视觉小说 URL: https://foreverse.app/zh/theater 书架上任何一本小说，阅读中途一键切成视觉小说演出：逐字台词、说话人名牌、AI 场景配图、剧情选择肢。不转换、不丢进度，选出来的走向真实写回书的分支。 ### Tavern Roleplay Chat with Character Cards URL: https://foreverse.app/roleplay Immersive character roleplay on mobile: import SillyTavern character cards, worldbooks, regex and Quick Reply automations, then chat with any BYOK model. ### 酒馆角色扮演聊天 — 角色卡 · 世界书 · 自动化 URL: https://foreverse.app/zh/roleplay 把 SillyTavern 酒馆装进手机：角色卡、世界书、正则脚本、Quick Reply 自动化与 STscript 全兼容，群聊与沉浸式背景开箱即用。 ### AI Companion with Real Memory — Lives on Your Phone URL: https://foreverse.app/companion A companion with a home, not a chat box: line-editable long-term memory, anniversaries they remember first, optional proactive messages with quiet hours, moment cards and selfies drawn from one confirmed reference set — data stored on your device. ### AI 伴侣 — 有记忆、会主动、住在你手机里 URL: https://foreverse.app/zh/companion 不是聊天框里的空壳：长期记忆逐条可查可改、纪念日 TA 先记得、主动消息可开可关、时刻卡与自拍按同一套确认过的参考图生成；关系数据全部存在你手机本地。 ### Multimodal Story Generation — Image, Voice, Video URL: https://foreverse.app/multimodal Generate illustrations, narration, and video from any selected passage or chat reply: Imagen, gpt-image, TTS voices, and Veo run inside the reading surface. ### 多模态生成 — 选区配图、配音朗读、配视频 URL: https://foreverse.app/zh/multimodal 为任意选区或聊天回复生成插画、朗读与视频：Imagen / gpt-image 配图、TTS 配音、Veo 配视频，全部嵌在阅读页与聊天页内完成。 ### AI Agent for Your Bookshelf — Reading, Cards, Automation URL: https://foreverse.app/agent Foreverse ships an in-app AI agent workbench on your phone: it reads your novels, completes worldbooks, organizes notes, and starts roleplay sessions from character cards — with a native tool-call loop, visible reasoning, @ references that jump between books and cards, and customizable skills. ### 智能体 Agent — 会读你书架、能干活的 AI 助手 URL: https://foreverse.app/zh/agent 新梦内置可干活的 AI 智能体：读小说、补世界书、整理笔记、用角色卡开新会话；原生工具调用真读真写，思考流与每步操作全程可见，@ 引用即可在小说和角色卡之间跳转，技能目录高度自定义。 ### Bring Your Own Key — 60+ AI Providers URL: https://foreverse.app/byok Connect your own API keys to 60+ providers: OpenAI, Anthropic, Gemini, DeepSeek, xAI and any compatible endpoint. Keys stay AES-encrypted on device. ### 自定义 API · BYOK — 自带 Key 接 60+ 模型供应商 URL: https://foreverse.app/zh/byok 把自己的 API Key 填进新梦：OpenAI、Anthropic、Gemini、DeepSeek、xAI 与任何兼容端点；密钥 AES 加密只存本机，请求直连供应商。 ### AI Audiobook from Any EPUB or TXT — Listen to Web Novels URL: https://foreverse.app/listen Turn any imported novel into an audiobook on your phone: system TTS for free, or online AI voices through your own key. Chunked chapter cache for commutes, 0.5–2.5x speed, sleep timer, and a reader that follows the narration. ### AI 听书 — 把任何 txt/epub 小说变成有声书 URL: https://foreverse.app/zh/listen 导入的小说直接听：系统 TTS 免费，在线 AI 音色走自己的 Key；整章分段缓存适合通勤，0.5–2.5 倍速、定时关闭，开着阅读器听页面跟着朗读走。 ### Community Character Card Showcase — 16 Picks with Covers, Lorebooks, Greetings URL: https://foreverse.app/cards Sixteen picks from the community hub's public character cards (90+ and growing), each with its own detail page: full personas, greetings, embedded lorebooks, example dialogue. AI-generated covers with origin marks; import any card from the app's Community tab. ### 社区精选角色卡 — 16 张带封面、世界书与开场白的公开卡 URL: https://foreverse.app/zh/cards 社区公开角色卡全量索引（90+ 张持续上新）+ 16 张精选：完整人设、开场白、随卡世界书、示例对话，每张卡有独立详情页。封面 AI 生成并带来源标记；在 App 社区页搜卡名一键导入开聊。 ### BYOK Provider Directory — 62 Presets, Every Modality Flagged URL: https://foreverse.app/byok/providers Every factory-preset BYOK provider in Foreverse, grouped and modality-flagged: frontier labs, mainland-China majors, aggregators, fast-inference clouds, enterprise platforms, local runtimes. Anything else connects as a custom OpenAI/Anthropic-compatible endpoint. ### BYOK 内置供应商目录 — 62 家预置逐家列明模态 URL: https://foreverse.app/zh/byok/providers Foreverse 出厂预置的 BYOK 供应商全列表：国际旗舰、中国大陆主流、聚合中转、推理加速云、企业平台、本地自托管分组列明，逐家标注文本/生图/视频/语音模态。不在列表的服务用自定义兼容端点接入。 ### Continue a Story with AI URL: https://foreverse.app/ai-story-continuation Use Foreverse to continue long fiction with AI while keeping alternate branches, reader context, and story notes organized. ### AI 续写小说 App — 把手机里的 txt 接着写下去，写给自己看 URL: https://foreverse.app/zh/ai-story-continuation 把手机里的 txt/epub 接着写下去：划词让 AI 续写，新内容落在分支上，原文一个字不改。315 本真实网文一次全选导入实测；62 家供应商自带 Key 直连，或官方渠道注册送 5000 积分约合 260 段续写。写给自己看，不用发表。 ### Branching Fiction Writer URL: https://foreverse.app/branching-fiction-writer Explore alternate scenes, endings, and rewrite paths without overwriting your main draft or reading progress. ### Story World Organizer URL: https://foreverse.app/story-world-organizer Keep character notes, chapter context, lore, and continuation prompts close to the reading surface. ### BYOK AI Writing Tool URL: https://foreverse.app/byok-ai-writing Bring your own AI provider key for private story work and choose the models that fit each reading or writing session. ### 小说开 if 线 — 在分支上重写剧情，原文一个字不动 URL: https://foreverse.app/zh/branching-fiction-writer 想改掉小说里某段剧情？在新梦从那一段直接分叉：改结局、救角色、换视角，每条 if 线都是独立分支，与原文并排保存、可对比可回退。「重新生成」也是分支，历史抉择卡随时 fork；本机 txt/epub 直接导入，每条分支还能换不同模型来写。 ### 小说设定管理 — 角色、世界书、剧情线贴着阅读面，AI 不忘设定 URL: https://foreverse.app/zh/story-world-organizer AI 续写老忘设定？整本塞给模型不是办法：窗口装不下，装下了注意力也散。新梦把设定拆成角色卡和世界书条目（chara_card_v3 兼容），续写哪段就只注入相关条目，阅读器内「本书 AI 资料」随读随改——结构化供给，不是硬塞原文。 ### BYOK 自带 Key 的 AI 写作 App — 62 家供应商，密钥只存你手机 URL: https://foreverse.app/zh/byok-ai-writing 不想为 AI 写作再订阅一个月费？新梦是 BYOK 写作 App：把 DeepSeek、智谱、月之暗面等 62 家供应商的 API Key 填进来，密钥系统级加密只存本机，请求直连供应商零加价——成本就是牌价本身。文本、生图、朗读、视频各设默认模型；不想配 Key 也有官方渠道，注册送 5000 积分约合 260 段续写。 ### SillyTavern on Android — No Termux, No Server, Cards Just Import URL: https://foreverse.app/sillytavern-android Run your tavern as a native Android app: chara_card_v3 cards (PNG/JSON/.charx), worldbooks with V3 decorators, regex scripts, Quick Reply v2, group chat, presets, beautify packs. No Termux session to keep alive — the screen locks and nothing dies. Import cards from files, the community hub, or a chub/JanitorAI URL; bring your own key across 62 providers or use the official metered channel. ### 手机上开酒馆 — 不用 Termux、不用部署，锁屏不断线 URL: https://foreverse.app/zh/sillytavern-android 原生 Android App 直接读酒馆那套格式：chara_card_v3 角色卡（PNG/JSON/.charx）、世界书（含 V3 decorators）、正则、Quick Reply v2、群聊、预设、美化包。没有要保活的 Termux 会话，锁屏回来接着聊；卡从本机文件、社区或 chub/JanitorAI 链接导入，自带 62 家供应商的 Key 或走官方按量渠道。 ### An AI Companion That Actually Remembers — Memory as Files You Can Read URL: https://foreverse.app/companion-with-memory A Foreverse companion keeps long-term memory as files on your phone: open the list, read every entry, edit what's wrong, delete what shouldn't stay. Anniversaries, journals, and the relationship archive live the same way, and the whole companion exports as one package when you switch phones. Memory you can audit, not a black box on someone's server. ### 有真记忆的 AI 伴侣 — 记忆存在你手机里，可以自己看、自己改 URL: https://foreverse.app/zh/companion-with-memory 新梦伴侣的长期记忆是你手机本机的文件：App 里逐条列出，看得到记了什么，记错了当场改，不想留的删掉。纪念日、日记、关系档案同样存在本机，换手机整包导出带走。TA 记得什么、记在哪里，两层你都查得到。 ### An AI That Texts You First — Memory-Driven, Not Scheduled URL: https://foreverse.app/companion-texts-first Foreverse companions send the first message because they remembered — the plan you mentioned, the anniversary you set — not because a campaign timer fired. Off by default, frequency you control, quiet hours, one master switch; memory lives in editable files on your phone. ### 会主动给你发消息的 AI — 因为想起你，不是定时推送 URL: https://foreverse.app/zh/companion-texts-first 新梦的 AI 伴侣会主动开口：消息从你们的聊天和 TA 的记忆里长出来——你提过的纪念日、说过的烦心事，不是运营定时群发。默认关闭、频率可调、有免打扰、一键全关；记忆存在你手机的文件里，可自己看、自己改。 ======================================================================== ## GLOSSARY (English) ### Character card URL: https://foreverse.app/docs/glossary#character-card A portable file holding an AI character's full definition: description, personality, scenario, greeting, example dialogue, plus optional embedded lorebook and regex scripts. Usually a PNG (data hidden in image metadata) or JSON file, interchangeable across apps. ### chara_card_v3 URL: https://foreverse.app/docs/glossary#chara-card-v3 The third-generation community spec for character cards: adds assets (sprites, audio), nickname, multilingual creator notes, group greetings, and lorebook decorators on top of v2's versioned envelope. Backward compatible with v2; the current mainstream format. ### .charx URL: https://foreverse.app/docs/glossary#charx The zip container defined by chara_card_v3: card JSON packed together with images, expression sprites, and audio as one distributable file. Suited to media-heavy cards. ### First message (first_mes / Greeting) URL: https://foreverse.app/docs/glossary#first-message The card-defined opening line, spoken by the character, that sets scene and tone. Strong greetings give a concrete situation and a hook to respond to rather than a generic hello. ### Alternate greetings URL: https://foreverse.app/docs/glossary#alternate-greetings Multiple switchable openings on one card, each a different starting scene or route. In the spec since v2; pick one when starting a chat. ### Persona URL: https://foreverse.app/docs/glossary#persona The definition of you in the conversation: name, identity, appearance, background. Where the character card describes the AI's side, the persona describes the user's. ### Lorebook (World Info) URL: https://foreverse.app/docs/glossary#lorebook An on-demand dossier system: each entry carries keywords, and when recent chat text matches them, the entry is injected into the prompt. It is not model memory — it is a mechanism that hands the model files about whatever got mentioned, keeping long-story lore consistent. ### Constant entry URL: https://foreverse.app/docs/glossary#constant-entry A lorebook entry injected every turn without keyword matching, for load-bearing world rules (magic system, era). Each one is rent paid every turn; keeping them in single digits is standard advice. ### Recursion URL: https://foreverse.app/docs/glossary#recursion Injected lorebook content triggering further entries. Deliberately writing related keywords inside an entry's text builds an association web: mention the sect, and the sect's heirloom sword rides along. ### Scan depth URL: https://foreverse.app/docs/glossary#scan-depth How many recent turns the lorebook scans for keyword matches. If a keyword last appeared beyond the window, the entry stays silent — the most common cause of 'works sometimes, not others.' ### Token budget URL: https://foreverse.app/docs/glossary#token-budget The per-turn cap on lorebook injection. Matched entries inject by priority until the budget fills; the rest drop silently. 'Lore stops working in long chats' usually means a few oversized entries ate the budget. ### V3 decorators URL: https://foreverse.app/docs/glossary#decorators @@-prefixed directives on chara_card_v3 lorebook entries (@@activate, @@dont_activate) for finer control of activation and injection. V2-only apps ignore them. ### Tavern URL: https://foreverse.app/docs/glossary#tavern Community shorthand for SillyTavern and compatible roleplay frontends, and by extension the whole import-a-card-and-chat playstyle. Core tavern capabilities: cards, lorebooks, regex, presets, group chat. ### Preset URL: https://foreverse.app/docs/glossary#preset A switchable conversation configuration: how the system prompt is organized, sampling parameters, context assembly order. Famous community presets are prompt-engineering know-how, packaged. ### Regex script URL: https://foreverse.app/docs/glossary#regex-script Find-and-replace rules applied to chat text, shipped with a card or global. Uses range from fixing a model's verbal tics to full beautification (turning plain-text markers into themed HTML). Scope can be display-only so fed-back history stays clean. ### Beautification URL: https://foreverse.app/docs/glossary#beautify The craft of rendering chat as themed interfaces with regex plus HTML/CSS: status bars, collapsible reasoning, clickable choices, whole chat skins. The chain: model outputs plain-text markers, regex replaces them with HTML, a sandboxed WebView renders it. ### Quick Reply URL: https://foreverse.app/docs/glossary#quick-reply Stored one-tap command buttons, often combined with STscript into automation: one button runs 'summarize the scene and archive it.' ### STscript URL: https://foreverse.app/docs/glossary#stscript SillyTavern's slash-command scripting language (/send, /setvar, ...) for chat automation. Foreverse supports its high-frequency subset. ### Macros ({{char}} / {{user}}) URL: https://foreverse.app/docs/glossary#macros Placeholder variables in prompts: {{char}} expands to the character's name, {{user}} to yours. They expand at prompt time only — not at display time. ### Swipe URL: https://foreverse.app/docs/glossary#swipe Multiple alternative replies for the same turn: swipe to regenerate while keeping earlier candidates switchable. The key difference from plain rerolling is that nothing gets destroyed. ### Group speaking strategy URL: https://foreverse.app/docs/glossary#group-chat-strategy The rule deciding who speaks next in a multi-character chat. Natural rotation arbitrates in three tiers — mentions first, talkativeness-weighted draw, list-order fallback — with list order, manual, and random as alternatives. ### Talkativeness URL: https://foreverse.app/docs/glossary#talkativeness A standard card field weighting how eagerly a character volunteers lines in group chat. High for extroverts who grab scenes, low for the laconic. ### BYOK (Bring Your Own Key) URL: https://foreverse.app/docs/glossary#byok Connecting your own API key directly to a model provider. The test: the key lives only on your device, requests hit the official endpoint, and prices are the provider's list prices. The alternative is platform proxying, where your chats pass through someone's server. ### OpenAI-compatible endpoint URL: https://foreverse.app/docs/glossary#compatible-endpoint A third-party service implementing the OpenAI API protocol. A custom API host plus key connects any compatible service; the host must include its full version prefix (such as /v1). ### Prompt caching (KV cache) URL: https://foreverse.app/docs/glossary#prompt-caching Provider discounts for repeated prefixes: the part of a request identical to recent requests bills at cache rates (commonly 10-50% of base). Long chats resend full history every turn, so hit rates decide the bill's order of magnitude. ### Context window URL: https://foreverse.app/docs/glossary#context-window The token ceiling a model handles per request — persona, lorebook, history, and your message all share it. When it overflows, old history gets truncated or summarized: the direct cause of 'long chats forget.' ### Thinking budget URL: https://foreverse.app/docs/glossary#reasoning-budget The internal reasoning tokens a reasoning model spends before answering, billed as output. Any output cap for such models must include this budget, or the visible reply can come back empty. ### Branch URL: https://foreverse.app/docs/glossary#branch The structure AI continuations land in: each continuation attaches beside the original text, which is never overwritten; multiple directions coexist, comparable and continuable. Version control for narrative — rerolling destroys, branching keeps. ### VN Theater URL: https://foreverse.app/docs/glossary#vn-theater A reading mode that performs the novel you are reading as a visual novel in place: typewriter dialogue, speaker nameplates, story choices that pause for you — choices write back as real branches. Entry and exit share one reading position; no converted project is created. ### Speaker attribution URL: https://foreverse.app/docs/glossary#speaker-attribution Deciding which character a quoted line belongs to. Foreverse's stance: attribute only on strong structural evidence and send everything uncertain to the narrator — misattribution costs more than abstention. ### Moving box URL: https://foreverse.app/docs/glossary#moving-box The whole-package export for AI companion data: persona, individual memories, stickers, avatar, and transcripts in one zip, restorable on another device. Import validates against a whitelist and rolls back on failure. ======================================================================== ## 术语表（中文） ### 角色卡（Character Card） URL: https://foreverse.app/zh/docs/glossary#character-card 承载一个 AI 角色全部定义的可携带文件：人设描述、性格、场景、开场白、示例对话，以及可选的内嵌世界书和正则脚本。常见形态是一张 PNG 图片（数据藏在图片元数据里）或 JSON 文件，跨应用通用。 ### chara_card_v3 URL: https://foreverse.app/zh/docs/glossary#chara-card-v3 角色卡格式的第三代社区规范：在 v2 的版本化信封之上增加资产（立绘、音频）、nickname、多语言创作者备注、群聊开场白和世界书 decorators，向下兼容 v2。是当前创作和分发的主流格式。 ### .charx URL: https://foreverse.app/zh/docs/glossary#charx chara_card_v3 定义的 zip 容器格式：卡片 JSON 与图片、表情立绘、音频等资产打包成一个文件分发。适合携带媒体资源较多的角色卡。 ### 开场白（first_mes / Greeting） URL: https://foreverse.app/zh/docs/glossary#first-message 角色卡里定义的第一条消息，由角色说出，为对话定下场景和语气。质量高的开场白通常给出具体情境和可回应的钩子，而不是泛泛的问候。 ### 备选开场白（Alternate Greetings） URL: https://foreverse.app/zh/docs/glossary#alternate-greetings 同一张卡的多个可切换开场白，对应不同的起始场景或剧情线。v2 起进入规范；导入后可以在开聊时选择用哪一条开场。 ### Persona（用户人设） URL: https://foreverse.app/zh/docs/glossary#persona 「你」在对话中的角色设定：名字、身份、外貌、背景。与角色卡相对，persona 描述的是用户侧。模型据此称呼你、理解你在剧情中的位置。 ### 世界书（World Info / Lorebook） URL: https://foreverse.app/zh/docs/glossary#lorebook 按需注入的设定档案库：每个条目带关键词，对话文本命中关键词时该条目被注入到发给模型的提示词里。它不是模型的记忆，而是一套「谁被提到就递谁档案」的机制，用来维持长篇设定一致。 ### 常驻条目（Constant Entry） URL: https://foreverse.app/zh/docs/glossary#constant-entry 不做关键词匹配、每轮都注入的世界书条目，适合世界观根基（魔法体系、时代背景）。每一条都是每轮必付的 token 成本，通常建议控制在个位数。 ### 递归激活（Recursion） URL: https://foreverse.app/zh/docs/glossary#recursion 被注入的世界书条目内容本身再触发其它条目的机制。在条目正文里主动写出相关条目的关键词，可以让设定形成关联网：提到门派时，门派的镇派之宝档案自动跟上。 ### 扫描深度（Scan Depth） URL: https://foreverse.app/zh/docs/glossary#scan-depth 世界书做关键词匹配时回看的对话轮数。关键词只在更早的楼层出现过、超出扫描窗口，条目就不会触发——这是「设定时灵时不灵」最常见的原因。 ### token 预算 URL: https://foreverse.app/zh/docs/glossary#token-budget 世界书单轮注入量的上限。命中的条目按优先级注入，预算用满后低优先级条目被静默丢弃——长对话里「设定聊着聊着失效」通常是预算被少数超长条目吃光。 ### V3 decorators URL: https://foreverse.app/zh/docs/glossary#decorators chara_card_v3 世界书条目上以 @@ 开头的指令（如 @@activate、@@dont_activate），用于更精细地控制条目的激活条件和注入位置。只支持 v2 的应用会忽略它们。 ### 酒馆 URL: https://foreverse.app/zh/docs/glossary#tavern 中文社区对 SillyTavern 及同类角色扮演前端的统称，也泛指「导入角色卡与模型对话」这一玩法。酒馆类应用的核心能力：角色卡、世界书、正则、预设、群聊。 ### 预设（Preset） URL: https://foreverse.app/zh/docs/glossary#preset 一套可切换的对话配置：系统提示词的组织方式、采样参数（温度等）、上下文组装顺序。社区流传的知名预设本质是提示词工程成果的打包分发。 ### 正则脚本（Regex Script） URL: https://foreverse.app/zh/docs/glossary#regex-script 对聊天文本做查找替换的规则，随卡分发或全局生效。用途从修正模型口癖到整卡美化（把纯文本标记替换成主题化 HTML）。作用域可控制只影响显示、不改写回喂给模型的历史。 ### 美化 URL: https://foreverse.app/zh/docs/glossary#beautify 用正则加 HTML/CSS 把聊天渲染成主题化界面的创作手艺：状态栏、可折叠思维链、可点选项、整套聊天皮肤。链路是模型输出纯文本标记、正则替换成 HTML、沙箱 WebView 渲染。 ### Quick Reply URL: https://foreverse.app/zh/docs/glossary#quick-reply 一键发送的预存指令按钮，常与 STscript 组合成自动化：点一个按钮完成「总结当前剧情并存档」这类多步操作。 ### STscript URL: https://foreverse.app/zh/docs/glossary#stscript SillyTavern 的斜杠命令脚本语言（/send、/setvar 等），用于编排聊天自动化。Foreverse 支持其高频子集。 ### 宏（{{char}} / {{user}}） URL: https://foreverse.app/zh/docs/glossary#macros 提示词里的占位变量：{{char}} 展开为角色名、{{user}} 展开为用户名。注意宏只在发给模型的 prompt 期展开，显示期不展开。 ### Swipe URL: https://foreverse.app/zh/docs/glossary#swipe 同一轮对话的多个备选回复：不满意当前回复时左右滑动重新生成，候选并存可切换。与「重新生成覆盖旧回复」的关键区别是旧候选不丢。 ### 群聊发言策略 URL: https://foreverse.app/zh/docs/glossary#group-chat-strategy 多角色群聊里决定「下一句谁说」的规则。自然轮替按「点名优先、话痨值加权、顺序兜底」三层裁决；另有列表顺序、手动指定、随机抽选三种策略可切换。 ### 话痨值（Talkativeness） URL: https://foreverse.app/zh/docs/glossary#talkativeness 角色卡上的标准字段，表示该角色在群聊里主动发言的倾向权重。外向角色设高值多抢戏，寡言角色设低值偶尔开口。 ### BYOK（Bring Your Own Key） URL: https://foreverse.app/zh/docs/glossary#byok 自带 API 密钥直连模型供应商的接入方式。判别标准：密钥只存在你设备上、请求直连官方端点、价格是供应商原价。与之相对的是平台代理模式（你的对话经过平台服务器）。 ### OpenAI 兼容端点 URL: https://foreverse.app/zh/docs/glossary#compatible-endpoint 实现了 OpenAI API 协议的第三方服务地址。填自定义 API 地址加密钥即可接入任何兼容服务，注意地址要含完整版本前缀（如 /v1）。 ### Prompt 缓存（KV Cache） URL: https://foreverse.app/zh/docs/glossary#prompt-caching 供应商对重复前缀的计费折扣机制：请求开头与近期请求相同的部分按缓存价计费（常见为原价一到五折）。长对话每轮重发全部历史，缓存命中与否直接决定账单量级。 ### 上下文窗口 URL: https://foreverse.app/zh/docs/glossary#context-window 模型单次请求能处理的 token 上限，人设、世界书、历史、你的消息都要装进这个窗口。窗口装不下时旧历史被截断或摘要，这是「聊久了忘事」的直接原因。 ### 思考预算 URL: https://foreverse.app/zh/docs/glossary#reasoning-budget 推理型模型在给出答案前产出的内部思考 token，按输出计费。给这类模型设输出上限时必须把思考预算算进去，否则正文可能一个字都不剩。 ### 分支（Branch） URL: https://foreverse.app/zh/docs/glossary#branch AI 续写的组织结构：每次续写作为独立分支挂在原文旁，原文永不被覆盖，多个走向并存可比较、可继续。类比代码的版本管理，reroll 是覆盖，分支是保留。 ### 剧场模式（VN Theater） URL: https://foreverse.app/zh/docs/glossary#vn-theater 把正在读的小说就地切换成视觉小说演出的阅读形态：台词逐字上屏、说话人有名牌、剧情分岔停下等你选，选择以分支写回原书。进出共用同一阅读进度，不产生新工程。 ### 说话人归因 URL: https://foreverse.app/zh/docs/glossary#speaker-attribution 判断一句引语出自哪个角色的技术问题。Foreverse 的立场是只认结构上有把握的证据、拿不准一律进旁白——错误归因比漏归因伤害大。 ### 搬家包 URL: https://foreverse.app/zh/docs/glossary#moving-box AI 伴侣数据的整包导出格式：人设、逐条记忆、贴纸、头像、聊天转录打成一个 zip，换设备导入还原。导入侧有白名单校验和失败回滚。 ======================================================================== ## COMPARISONS — verdicts and FAQs ### Foreverse vs RikkaHub: Open-Source LLM Client, or Tavern + Reader in One? URL: https://foreverse.app/compare/foreverse-vs-rikkahub Updated: 2026-07-14 Verdict: RikkaHub is one of the most polished native LLM clients on Android: Material You, MCP, Mermaid/LaTeX rendering, message branching, a clean BYOK setup, source published on GitHub, and since late 2025 lorebook import/export too. Foreverse bets on a different layer of depth: card-embedded assets that activate with the card, QR and STscript-style automation, a long-fiction reader, and a companion with managed memory. If you want a general-purpose AI chat client, RikkaHub is genuinely good; if you want the tavern, the bookshelf, and the relationship in one pocket, come here. Q: Both import character cards — where is the difference? A: In the card-embedded assets. After import, RikkaHub needs the embedded lorebook imported separately and attached to the assistant manually (its community bridge tool documents this); Foreverse activates embedded lorebooks with the card per the chara_card_v3 spec, runs regex scripts in rendering, supports alternate greetings, and imports group-chat JSON. The more complex the card, the wider the gap. Q: Is RikkaHub free and open source? What about Foreverse? A: RikkaHub is free to use with source published on GitHub, but its license carries commercial-use restrictions, so it is not OSI open source in the strict sense — check its LICENSE for details. Foreverse is closed source with a free core; model usage goes through your own API keys at provider list price, plus an optional official metered channel. Q: Is migrating from RikkaHub to Foreverse easy? A: Character cards (PNG/JSON) and lorebook JSON files work directly. Provider configs need re-entering your keys (a few minutes). Chat histories use different formats on each side; there is no one-click migration. ### Foreverse vs SillyTavern: Which Tavern Fits Your Phone in 2026? URL: https://foreverse.app/compare/foreverse-vs-sillytavern Updated: 2026-06-08 Verdict: SillyTavern remains the most powerful roleplay frontend ever built — on a desktop. Foreverse is for the hours you are not at that desk: a native Android app that imports the same chara_card_v3 cards, lorebooks, and regex scripts, adds a long-fiction reader, and needs no server, Termux, or git pull. Many of our users run both. Q: Can I import my SillyTavern characters into Foreverse? A: Yes. PNG and JSON character cards (v1/v2/v3), lorebook JSON files, regex scripts, and Quick Reply sets all import. Embedded character books come along with their cards. Q: Is Foreverse a fork of SillyTavern? A: No. Foreverse is a native Android app built from scratch that implements compatibility with SillyTavern's data formats and core behaviors. No SillyTavern code runs inside it. Q: Do I still need a PC or server for Foreverse? A: No. The app talks directly to model providers with your own API keys. There is nothing to deploy, update, or keep awake. Q: Should I switch if SillyTavern already works for me? A: If your setup is stable and desktop-bound, keep it — genuinely. Add Foreverse when you want the same characters available on your phone without running infrastructure. ### Foreverse vs Character.AI: Own Your Characters or Rent Them? URL: https://foreverse.app/compare/foreverse-vs-character-ai Updated: 2026-06-08 Verdict: Character.AI is the easiest way to start AI roleplay: open the app, pick from millions of characters, chat free. The price is control — its models only, its content filter, and characters that cannot leave the platform. Foreverse trades that convenience for ownership: portable character cards, 60+ model providers via your own keys, and a library that lives on your phone, not in someone's database. Q: Can I move my Character.AI personas into Foreverse? A: There is no official export from Character.AI. The common route is recreating the persona as a chara_card_v3 card — description, personality, greeting, example dialogue. Community tools can help reconstruct definitions you wrote yourself. Q: Is Foreverse free like Character.AI? A: The app's core is free. Model usage goes through your own API keys at provider list prices, or through optional credits for official channels. There is no subscription wall in front of your own keys. Q: Does Foreverse have a content filter? A: Foreverse does not add a platform-wide filter on top of your BYOK traffic; the policies that apply are those of the model provider you choose. You pick the provider whose terms fit your writing. ### Foreverse vs NovelAI: Subscription Storyteller or BYOK Reader-Writer? URL: https://foreverse.app/compare/foreverse-vs-novelai Updated: 2026-06-08 Verdict: NovelAI bets on vertical integration: in-house models tuned for prose, a lorebook editor, anime-grade image generation, all under one subscription. Foreverse bets on openness: your epub/txt library, branching continuation on top of it, and whichever frontier model you want via your own keys. If you write in NovelAI's editor daily, stay. If you read long fiction and want models to come to your books, that is us. Q: Can Foreverse match NovelAI's prose quality? A: Prose quality follows the model you bring. Frontier general models in 2026 write strong fiction with good prompting; NovelAI's edge is consistency of a house style. With BYOK you trade a curated voice for the freedom to chase whichever model currently writes best. Q: Does Foreverse have anime image generation like NovelAI? A: Foreverse generates scene illustrations through providers like Google Imagen or gpt-image via your keys. For specifically anime-styled output, NovelAI's dedicated diffusion models remain the stronger pick. Q: Can I import my NovelAI stories? A: Export your stories as text and import them as books in Foreverse; lorebook entries can be recreated in the worldbook editor. There is no one-click converter today. ### Foreverse vs AI Dungeon: Guided Adventures or Your Own Library? URL: https://foreverse.app/compare/foreverse-vs-ai-dungeon Updated: 2026-06-08 Verdict: AI Dungeon is a game first: scenarios, quests, multiplayer parties, a marketplace of community adventures. Foreverse is a library first: your novels and characters, with branching continuation and SillyTavern-compatible cards. If you want to drop into a pre-built dungeon tonight, play AI Dungeon. If you want your own worlds to grow and stay yours — on your phone, on your keys — that is Foreverse. Q: Is Foreverse a text adventure game? A: Not in the quest sense. Foreverse is a reader and tavern: continuation, branching, and character chat on top of your own library. You can absolutely run adventure-style scenes, but there is no scripted quest marketplace. Q: Can I recreate my AI Dungeon scenarios in Foreverse? A: Export your story text and import it as a book; recurring world rules become worldbook entries and characters become cards. Scenario-specific mechanics like stat tracking would need Quick Reply automation to emulate. Q: Which is cheaper? A: Depends on volume. AI Dungeon's free tier is generous for casual play; heavy use pushes you to subscriptions. Foreverse's core is free and model usage rides your own API pricing — heavy users usually pay less at list price, light users may not. ### Foreverse vs JanitorAI: Community Scale, or Cards as Files You Own? URL: https://foreverse.app/compare/foreverse-vs-janitor-ai Updated: 2026-07-17 Verdict: JanitorAI earned its place: a giant card community, a free built-in model (JLLM), official proxy support so you were never locked to one LLM, and a beta mobile app since early 2026. What it does not give you is the card as a file — no official export that we could find. Foreverse does, and the move is one paste: as of 2026-07 a JanitorAI character link imports directly, joined by a novel reader, audiobook listening, and an AI companion. Q: Can I move my JanitorAI characters into Foreverse? A: Yes, and it is the easiest migration on this site: paste the character's page link (jannyai mirror links also work) into Foreverse's card import, and it arrives as a chara_card_v3 file you own. Some network environments get blocked by JanitorAI's download protection — if that happens, retry on a different network. Chat histories are separate; there is no one-click chat migration. Q: Does Foreverse have a free model like JLLM? A: Not a house model — Foreverse does not train its own LLM. Instead, the app core is free with no daily message caps, signing up grants 5,000 credits on the official metered channel (1 credit = $0.0001, roughly 260 continuations), and BYOK connects 62 providers at list price with keys stored encrypted on your device. Q: Will my NSFW cards still work? A: The cards import regardless of content. Foreverse does not add a platform filter on top of your BYOK traffic; what applies is the content policy of the model provider you choose. Pick the provider whose terms fit what you write. ### Foreverse vs Talkie: Gacha Companion Platform, or a Companion You Own? URL: https://foreverse.app/compare/foreverse-vs-talkie Updated: 2026-07-17 Verdict: Talkie is MiniMax's overseas companion app, sibling of China's Xingye: Sensor Tower ranked it the fourth most-downloaded AI app in the US in the first half of 2024, on expressive voice and a gacha card layer its rivals don't have. It is also a closed loop: in-house models, platform moderation, and no character export we could find as of 2026-07. Foreverse trades that polish for ownership — cards as open files, 62 providers on your keys, chats on your device. Q: Can I move my Talkie characters into Foreverse? A: We found no official character export in Talkie as of 2026-07. The working route is a rebuild: write the persona, speaking style, and your relationship history into a chara_card_v3 card, and move durable facts into lorebook entries so they stay active in long chats. Our blog has a full migration walkthrough (the Character.AI version — the method transfers). Q: Does Foreverse have voice like Talkie? A: Yes, via BYOK: connect any OpenAI-compatible TTS provider — MiniMax's own open platform included — and give the character a voice, at the provider's list price on your key. What Foreverse does not have is Talkie's produced voice-call experience; that polish is real, and it is theirs. Q: What does Foreverse cost compared to Talkie's free tier? A: The app core is free with no ads and no message caps. Model usage runs on your own API keys at provider list price, or on the official metered channel (1 credit = $0.0001; signing up grants 5,000 credits, roughly 260 continuations). There is no gacha currency and nothing to pull. ### Foreverse vs StoryLord: A Web Library, or Books That Live on Your Phone? URL: https://foreverse.app/compare/foreverse-vs-storylord Updated: 2026-07-17 Verdict: StoryLord (gennovel.app) overlaps with Foreverse more than anything else on this list: upload TXT/DOCX/Markdown or pick a public-domain classic, steer the AI's continuation or rewrite, publish the result to a community library — 100 free credits at signup, one-time packs after. Which model writes for you is not disclosed as of 2026-07. Foreverse is the on-device answer: a native Android reader, 62 providers on your keys, and continuations that grow as branches beside the original, never over it. Q: Can I move my StoryLord continuations into Foreverse? A: Through text, yes: get the original plus generated chapters out as plain text (copy or whatever export the site offers at the time), save as a txt, and import it as a book in Foreverse to keep writing. There is no one-click migrator. Recurring world facts are worth re-entering as worldbook entries so long continuations stay consistent. Q: Which models does each side use? A: As of 2026-07 we could not find model names or a model picker anywhere on StoryLord's site — the platform decides. Foreverse is transparent by construction: 62 providers via your own keys at provider list price (keys encrypted on-device), or the official metered channel at 1 credit = $0.0001 with a 5,000-credit signup grant. Q: Does Foreverse have StoryLord's outline-first workflow? A: No, and credit where due — the approve-an-outline-then-generate loop is a genuinely good production process for whole books. Foreverse is built around the reading moment instead: select a passage, optionally ask for a few plot directions, and grow the one you like as a branch. Book-factory workflow: theirs. Read-and-continue: ours. ### Foreverse vs Sudowrite: The Author's Desk, or the Reader's Armchair? URL: https://foreverse.app/compare/foreverse-vs-sudowrite Updated: 2026-07-17 Verdict: Sudowrite is the most polished AI suite for authors drafting books to publish: Story Bible, the fiction-trained Muse model, 20+ prose tools, from $10/month billed annually ($19 monthly) as of 2026-07. Foreverse points the same technology the other way — import the published novel you are reading and grow continuations as branches beside it, free core, your own keys. Many people searching for a Sudowrite alternative actually need this other species. Writing a book: stay there. Reading one: come here. Q: Can I move my Sudowrite project into Foreverse? A: The manuscript, yes: export it as text and import it as a book — continuation and branching work on it immediately. Story Bible entries have no importer; durable facts (characters, world rules) are worth re-entering as worldbook entries, which trigger by keyword during continuation. If you are mid-draft as an author, though, staying in Sudowrite is usually the right call. Q: Is Foreverse a writing tool like Sudowrite? A: Not in the manuscript sense. There is no outline board, no revision workflow, no export-to-publisher path. Foreverse is a reader that writes: continuation, rewriting a passage on a branch, scene illustration, and audiobook narration on top of books you import. People do write original fiction in it, but the center of gravity is reading. Q: Which costs less? A: For drafting a whole novel, Sudowrite's Professional tier at $22/mo annual is genuinely competitive — usage included. For continuing scenes as a reader, metered wins: Foreverse's app is free, and a few continuations a week on your own DeepSeek or Gemini key costs pocket change. The crossover point is your monthly word count; do the arithmetic with real numbers. ### Foreverse vs NovelCrafter: Two BYOK Answers to Two Different Questions URL: https://foreverse.app/compare/foreverse-vs-novelcrafter Updated: 2026-07-17 Verdict: NovelCrafter and Foreverse argue the same side of the BYOK debate: your key, provider list price, no model lock-in. What differs is the room. NovelCrafter is a web writing platform — Codex story bible, series planning, team collaboration — at $4–20/month as of 2026-07, AI features from the $8 Hobbyist tier. Foreverse is a free native Android reader that grows continuations as branches beside books you import. Drafting a novel: theirs. Continuing one you are reading: ours. Q: Can I move my NovelCrafter novel into Foreverse? A: The prose, yes: export the manuscript as text and import it as a book — continuation and branching work immediately. Codex entries have no importer; re-enter the durable ones as worldbook entries (they trigger by keyword during continuation). Chat and revision history stay behind. NovelCrafter's post-cancel export policy makes getting your text out straightforward, which we note with respect. Q: Can I use the same API key for both? A: Yes, and that is the quiet advantage of both being BYOK: the OpenRouter or OpenAI or Anthropic key you configured for NovelCrafter works in Foreverse unchanged — 62 providers are preset, plus any OpenAI-compatible endpoint. Keys stay encrypted on your device and calls go direct at list price. Running both tools on one key during a trial month is a perfectly sane way to compare. Q: What does NovelCrafter's subscription buy that Foreverse doesn't charge for? A: Different goods. NovelCrafter's $4–20/mo buys manuscript infrastructure: the Codex, series planning, review tools, team seats. Foreverse's core is free because the product is a reader — continuation, branching, TTS listening, roleplay cards — and the only money that moves is your model usage: your keys at provider price, or the official metered channel (1 credit = $0.0001, 5,000 granted at signup). ### Foreverse vs Kindroid: The Deepest Companion, or a Companion You Can Open in a File Manager? URL: https://foreverse.app/compare/foreverse-vs-kindroid Updated: 2026-07-17 Verdict: Kindroid is the deepest pure companion app here: infinite long-term memory consolidation for every account, voice and video calls, contextual selfies, group chats — from $13.99/month on web as of 2026-07, memory add-ons stacking to $98.97. It is also a closed loop: in-house models only, no official export we could find. Foreverse trades that depth for custody — memory as files you can read and edit on your phone, 62 providers on your keys, and a novel reader in the same app. Q: Can I move my Kindroid companion into Foreverse? A: There is no official Kindroid export as of 2026-07. The community KinX tool (GitHub) can pull your Kin's backstory, key memories, journals, and chats to local JSON using your own auth token; from there it is a rebuild — backstory becomes the persona in a chara_card_v3 card, durable facts become memory entries or worldbook lines. The bond re-forms faster than you expect once the facts survive. Q: Does Foreverse have selfies and voice like Kindroid? A: Partly, and honestly: the companion can send photos generated through your own image providers, with a daily cap you control, and speaks via any OpenAI-compatible TTS on your key. There are no video calls and no dedicated selfie pipeline — Kindroid's media stack is genuinely ahead there. Foreverse's trade is that every photo and every memory ends up as a file you keep. Q: What does Foreverse cost next to Kindroid's subscription? A: Kindroid's full experience starts at $13.99/month on web ($15.99 in-app), with Ultra and MAX add-ons stacking to $98.97/month for maximum memory. Foreverse's core is free with no message caps; you pay only model usage — your own keys at provider list price, or the official metered channel (1 credit = $0.0001, 5,000 credits granted at signup). Light companion use costs pocket change; heavy use scales with your chosen model, visibly. ### Foreverse 对比 RikkaHub：开源 LLM 客户端，还是酒馆+阅读器一体？ URL: https://foreverse.app/zh/compare/foreverse-vs-rikkahub Updated: 2026-07-14 Verdict: RikkaHub 是 Android 上完成度最高的原生 LLM 客户端之一：Material You、MCP、Mermaid/LaTeX 渲染、消息分支，BYOK 配置体验干净利落，源码公开可读，2025 年底也加入了世界书的导入导出。Foreverse 押注的是另一层深度：卡内嵌资产随卡生效、QR 与 STscript 风格自动化、长篇小说阅读器、有记忆的伴侣。想要一个通用 AI 聊天客户端，RikkaHub 很好；想把酒馆、书架和关系装进同一部手机，来我们这。 Q: RikkaHub 和 Foreverse 都支持角色卡导入，差别在哪？ A: 差在随卡资产。RikkaHub 导卡后，卡内嵌的世界书要单独导入再手动关联助手（其社区桥接工具的文档明确说明了这一点）；Foreverse 按 chara_card_v3 规范让内嵌世界书随卡触发、正则脚本参与渲染、备选开场白可切换、群聊 JSON 可导入。卡越复杂，差距越明显。 Q: RikkaHub 是免费开源的吗？Foreverse 呢？ A: RikkaHub 免费下载、源码公开，但许可证对商业用途有专门条款，严格意义上不是 OSI 定义的开源，细节以其仓库 LICENSE 为准。Foreverse 闭源、核心功能免费，模型用量走你自己的 API Key（供应商原价直连），也有可选的官方渠道按量计费。 Q: 从 RikkaHub 迁移到 Foreverse 方便吗？ A: 角色卡（PNG/JSON）和世界书 JSON 直接复用。API 供应商配置需要重新填一遍 Key（几分钟）。聊天记录两边格式不同，目前没有一键迁移。 ### Foreverse 对比星野：平台养成，还是角色所有权？ URL: https://foreverse.app/zh/compare/foreverse-vs-xingye Updated: 2026-07-14 Verdict: 星野是国内 AI 陪伴赛道打磨最久的产品之一：MiniMax 的语音技术、写真卡面、关系档案、数百万月活的社区，拿来即玩的完成度很高。作为平台产品，它的模型体系是自家的，角色没有官方导出入口（截至 2026-07 我们未找到），重度使用会碰到免费额度类限制。Foreverse 用十分钟配置换回三样东西：模型自选（60+ 供应商）、角色是你手里的标准格式文件、聊天记录存在你设备上。两条路线服务两种玩家，这页把差异摊开。 Q: 星野的角色能搬到 Foreverse 吗？ A: 截至 2026 年 7 月，我们没有在星野里找到官方的角色导出入口。可行做法是重建：把角色设定、说话风格和你们的关系要点按 chara_card_v3 重写成卡，关系史整理成世界书常驻条目。我们博客有一篇完整的迁移过程记录（C.AI 版本，方法通用）。 Q: Foreverse 有星野那样的语音吗？ A: 有，走 BYOK：接任何 OpenAI 兼容语音端点（阿里云百炼、MiniMax 开放平台、xAI 等）给角色配音色，费用按供应商定价走你的 Key。体验取决于你接的供应商——事实上你可以接 MiniMax 自己的音色。 Q: Foreverse 怎么收费？会有体力类限制吗？ A: App 核心免费，没有按条数限速的设计。模型调用走你自己的 API Key 时按供应商原价直连、我们不加价；不想配 Key 可以用官方渠道按量计费（1 credit = $0.0001）。 ### Foreverse 对比猫箱：大厂内容流，还是自己的书房？ URL: https://foreverse.app/zh/compare/foreverse-vs-maoxiang Updated: 2026-07-14 Verdict: 猫箱是字节跳动在 AI 陪伴上的答卷，最强的地方非常字节：探索页像刷短视频一样刷角色、滑到即聊、心动指令卡按节日上新、语音自然。据报道它在 2024 年末仍保持增长，是赛道里的少数派。作为平台产品，模型体系是字节自家的、角色没有官方导出入口（截至 2026-07 我们未找到）、内容口径跟着平台审核走。Foreverse 押注相反的东西——你自己的模型、自己的角色文件、自己的书架。刷角色的爽感我们给不了，角色文件握在自己手里这件事，我们做得很足。 Q: 猫箱的角色能搬到 Foreverse 吗？ A: 截至 2026 年 7 月，我们没有在猫箱里找到官方的角色导出入口。做法同其它平台：把人设、说话风格、关系要点重建成 chara_card_v3 卡片，关系史写进世界书。重建后的角色七分像，剩下三分靠新的相处长回来。 Q: Foreverse 有猫箱那种「刷到即聊」的发现体验吗？ A: 没有那个量级。我们有社区卡站（浏览、详情预览、一键导入开聊），但内容池远小于字节系产品。发现体验是我们诚实承认的差距，所有权是我们的交换条件。 Q: 两个都免费，成本差在哪？ A: 猫箱免费背后是平台模型与平台数据；Foreverse 核心免费，模型走你自己的 API Key 按量原价（轻度用户一个月几块钱量级），或官方渠道按量计费。差别不在多少钱，在钱换来的控制权。 ### Foreverse 对比 SillyTavern：2026 年手机酒馆怎么选？ URL: https://foreverse.app/zh/compare/foreverse-vs-sillytavern Updated: 2026-06-08 Verdict: SillyTavern 依然是有史以来最强大的角色扮演前端——前提是你坐在电脑前。Foreverse 负责你不在书桌前的那些时间：原生安卓应用，导入同样的 chara_card_v3 角色卡、世界书和正则脚本，外加一个长篇小说阅读器，不需要服务器、Termux 和 git pull。我们的不少用户两边都在用。 Q: SillyTavern 的角色卡能直接导进 Foreverse 吗？ A: 能。PNG 和 JSON 角色卡（v1/v2/v3）、世界书 JSON、正则脚本、Quick Reply 集合都支持导入，卡内嵌的角色书会随卡一起进来。 Q: Foreverse 是 SillyTavern 的分支（fork）吗？ A: 不是。Foreverse 是从零开发的原生安卓应用，实现了对 SillyTavern 数据格式和核心行为的兼容，里面不运行任何 SillyTavern 的代码。 Q: 用 Foreverse 还需要电脑或服务器吗？ A: 不需要。应用拿你自己的 API Key 直连模型供应商，没有任何需要部署、更新、保活的东西。 Q: 我的 SillyTavern 用得好好的，要换吗？ A: 桌面用得稳就别换，真心话。在你想让同一批角色出现在手机上、又不想跑基础设施的时候，再加一个 Foreverse 就够了。 ### Foreverse 对比 Character.AI：角色是自己的，还是租的？ URL: https://foreverse.app/zh/compare/foreverse-vs-character-ai Updated: 2026-06-08 Verdict: Character.AI 是开始 AI 角色聊天最容易的方式：打开就有百万角色，免费开聊。代价是控制权——只能用它家模型、套同一个过滤器、角色无法离开平台。Foreverse 用十分钟的上手成本换回所有权：角色是开放格式的卡片文件，模型走你自己的 Key 任选 60+ 家，聊天库存在你手机上而不是别人的数据库里。 Q: Character.AI 的角色能搬到 Foreverse 吗？ A: C.AI 没有官方导出。常见做法是把人设重建成 chara_card_v3 卡片：描述、性格、开场白、对话示例。自己写过的定义可以借助社区工具整理还原。 Q: Foreverse 像 Character.AI 一样免费吗？ A: 应用核心免费。模型用量走你自己的 API Key（供应商原价），或者用可选的官方积分。你自己的 Key 前面没有订阅墙。 Q: Foreverse 有内容过滤器吗？ A: Foreverse 不在你的 BYOK 流量上叠加平台级过滤器，生效的是你所选模型供应商自己的政策。哪家的条款适合你的创作，你就选哪家。 ### Foreverse 对比 NovelAI：订阅制写作套件，还是 BYOK 读写一体？ URL: https://foreverse.app/zh/compare/foreverse-vs-novelai Updated: 2026-06-08 Verdict: NovelAI 押注垂直整合：自家为叙事特调的模型、成熟的 Lorebook 编辑器、动漫级生图，一个订阅全包。Foreverse 押注开放：你的 epub/txt 书库、在书上长分支续写、模型用自己的 Key 接任何一家。每天泡在 NovelAI 编辑器里写作的人，留下；从「读已有的书」出发、想让模型来迁就书的人，来我们这。 Q: Foreverse 的文笔能赶上 NovelAI 吗？ A: 文笔跟着你带来的模型走。2026 年的旗舰通用模型配合好的提示词，长文写作能力已经很强；NovelAI 的优势在于稳定的「家传风格」。BYOK 换来的是随时追逐当下最能写的那个模型的自由。 Q: Foreverse 有 NovelAI 那种动漫生图吗？ A: Foreverse 通过你的 Key 调用 Google Imagen、gpt-image 这类供应商生成场景插画。如果明确要动漫画风，NovelAI 的专用扩散模型目前仍是更强的选择。 Q: NovelAI 里写的故事能搬过来吗？ A: 把故事导出为文本、在 Foreverse 里作为书导入即可；Lorebook 条目可以在世界书编辑器里重建。目前没有一键迁移器。 ### Foreverse 对比 AI Dungeon：进别人的副本，还是养自己的书？ URL: https://foreverse.app/zh/compare/foreverse-vs-ai-dungeon Updated: 2026-06-08 Verdict: AI Dungeon 首先是个游戏：剧本（scenario）、行动指令、多人小队、社区冒险市场。Foreverse 首先是个书房：你的小说和角色，配上分支续写和 SillyTavern 兼容的卡片体系。今晚就想进个现成副本爽一把，去 AI Dungeon；想让自己的世界慢慢长大、并且永远属于你——在你手机上、用你的 Key——来 Foreverse。 Q: Foreverse 是文字冒险游戏吗？ A: 不是任务意义上的。Foreverse 是阅读器加酒馆：在你自己的书库上续写、开分支、和角色聊天。冒险风格的戏完全可以跑，但没有脚本化的任务市场。 Q: AI Dungeon 里的剧本能在 Foreverse 重建吗？ A: 把故事文本导出、作为书导入；反复出现的世界规则做成世界书条目，角色做成卡片。属性追踪这类剧本机制，需要用 Quick Reply 自动化来模拟。 Q: 哪个更便宜？ A: 看用量。AI Dungeon 免费层对轻度玩家够用，重度玩家会被推向订阅；Foreverse 核心免费、模型按你自己的 API 原价计费——重度通常更省，轻度未必。 ### Foreverse 对比彩云小梦：三选一的轻快，还是整本书的长跑？ URL: https://foreverse.app/zh/compare/foreverse-vs-caiyun-xiaomeng Updated: 2026-07-17 Verdict: 彩云小梦是中文 AI 续写的开山产品：2021 年 9 月上线，「三条走向选一条」和「平行世界」这套玩法就是它教给整个品类的，而且 2026 年仍在更新，接入了 DeepSeek R1 和豆包。它的长处是轻快：开个头、卡文时接一把、当场看三种可能。Foreverse 服务另一种人：把手机里那本几百万字的 txt 导进来接着写，续写落在分支上，原文一个字不改，几条时间线并存可对比回退，模型用你自己的 Key 在 62 家里任选。快玩选小梦，长期养书来我们这。 Q: 彩云小梦里写的故事能搬到 Foreverse 吗？ A: 能，走文本这条路：把小梦里的正文复制出来存成 txt，在 Foreverse 作为书导入，就能接着写；世界设定和角色关系建议整理成世界书条目，长篇续写时按关键词自动激活。没有一键迁移器，但纯文本这条路是通的。 Q: Foreverse 有小梦那种三选一吗？ A: 有相近的玩法：续写时可以让 AI 先给出几条剧情走向，挑一条展开。差别在产物上——小梦选完即走，Foreverse 每条展开都落在自己的分支上，挑剩的走向不会消失，回头随时换一条世界线重来。 Q: 两边都有免费额度，差在哪？ A: 机制不同。小梦是每日免费额度，用完靠会员或字数加量包（价格以 App 内显示为准）；Foreverse 核心功能免费且不设条数限速，模型费用走你自己的 API Key 按供应商原价，或官方渠道按量计费（1 credit = $0.0001，注册送 5000，约 260 段续写）。轻度使用两边都便宜，重度养书时结构差异才显出来。 ### Foreverse 对比 StoryLord：网页图书馆，还是手机书房？ URL: https://foreverse.app/zh/compare/foreverse-vs-storylord Updated: 2026-07-17 Verdict: StoryLord（gennovel.app）是这份清单里功能面与我们重叠最多的产品：上传 TXT/DOCX/Markdown 或从内置公版书库选一本，AI 按你给的方向续写或改写，写完一键发布到社区图书馆；注册送 100 积分，积分包一次性买断不搞订阅。它是网页产品，生成用什么模型截至 2026-07 官网未见说明。Foreverse 把同一件事装进手机原生 App：本机 txt/epub 导入（315 本真机实测）、续写落在原文旁边的分支上、62 家供应商自带 Key。想把续作发出去给人读，选它；想长期私养一本书，来我们这。 Q: StoryLord 里生成的续写能搬到 Foreverse 吗？ A: 能，走文本这条路：把原著和续写章节的正文取出存成 txt（能否整本导出以站内实际功能为准，复制粘贴这条路总是通的），在 Foreverse 作为一本书导入接着写。没有一键迁移器。反复出现的世界设定建议顺手整理成世界书条目，长篇续写时按关键词自动激活。 Q: Foreverse 有 StoryLord 那种「先出大纲再写正文」的流程吗？ A: 没有整本大纲工作台，这点它更像一道成书工序，做得也确实好。我们的续写以阅读现场为中心：划词或从段尾发起，可以先让 AI 给几条剧情走向再展开，每条展开落在自己的分支上。想要图书馆式的整本生产流程，它更顺手；想在读到哪儿写到哪儿，我们更顺手。 Q: 两边的模型差在哪？ A: 截至 2026-07，StoryLord 官网与帮助中心未见模型型号说明或自选入口，生成用什么由平台决定。Foreverse 模型全透明：62 家供应商自带 Key 按供应商原价直连（Key 加密只存本机），或官方渠道按量计费（1 credit = $0.0001，注册送 5000）。 ======================================================================== ## BLOG — descriptions and FAQs ### We Counted Em Dashes in 30 Top Roleplay Cards. The “ChatGPT Hyphen” Isn't the Tell You Think URL: https://foreverse.app/blog/em-dash-is-not-an-ai-tell Published: 2026-07-21 · Author: Deng Binjie The em dash got branded the “ChatGPT hyphen” and writers started scrubbing it out of their own prose. We went to measure it before adding a dash rule to our English AI-flavor detector — and found that 8 of the 30 top-starred human roleplay cards use em dashes above the density that flags a Chinese text as machine-written, while 2 of 5 tested model families use none at all. Any dash threshold that catches AI also false-flags a quarter of the best human cards. The numbers, the method, and the rule we shipped instead. Q: Is the em dash proof that a text was written by ChatGPT? A: No. It's weak evidence at best and it false-flags heavily. In our July 2026 count, 8 of the 30 top-starred human-written English roleplay cards used em dashes above 3 per 1,000 words of narration — the density that reliably flags machine text in Chinese — with one human card hitting 20.6. Meanwhile two of the five current model families we sampled produced zero em dashes in their raw output. A punctuation mark that a quarter of the best human writers lean on and some models never touch cannot carry an accusation. Q: Why is the long dash an AI tell in Chinese but not in English? A: Because the human baselines differ. Chinese web-fiction and roleplay prose almost never uses the double-width dash —— so when a model produces it constantly, the deviation from the human norm is enormous, and in our Chinese detector's calibration it is the single most reliable machine fingerprint. English narrative prose — especially the asterisk-roleplay tradition — inherited heavy dash usage from published fiction. Same punctuation, different baseline, opposite verdict. AI tells are language-specific; porting them across languages without re-measuring produces false accusations. Q: What should you check instead of em dashes? A: Cluster density of stock phrases (in our calibration, top human card greetings average 0.07 slop-lexicon hits; raw model greetings average 0.22), the "not X, but Y" contrast construction, emotion-cocktail formulas ("a mix of anticipation and dread"), and fishing-line endings ("the choice is yours"). None of these is proof either — flagship models in single-shot mode now pass phrase checks clean — but they separate in the right direction, which the em dash does not. ### AI Roleplay Slop Words: The 95-Phrase Reference List (2026) URL: https://foreverse.app/blog/ai-roleplay-slop-words-list Published: 2026-07-21 · Author: Deng Binjie A maintained, sourced reference of the 95 phrases that mark AI-generated roleplay prose: shivers down spines, whispers barely above themselves, mischief-sparkling eyes, and the "not X, but Y" construction. Cross-checked from Sukino's Banned Tokens, the Antislop paper (arXiv 2510.15061 — "Elara" runs 85,513x over the human baseline), and EQ-Bench's slop score. Plus two numbers most lists miss: top human-written cards average slop hits too, and flagship models' single-shot output now passes word-list checks clean. Q: What are slop words in AI roleplay? A: Stock phrases that language models over-produce relative to human writers — "barely above a whisper," "eyes sparkling with mischief," "shivers down her spine," "ministrations." The over-production is measurable: the Antislop paper (arXiv 2510.15061) found the character name "Elara" appearing 85,513 times more often in one model's fiction output than in human writing, and the trigram "heart hammered ribs" at 1,192x. Individually the phrases are ordinary English; it's the density and clustering that reads as machine output. Q: Is one slop phrase proof that text is AI-generated? A: No. Humans wrote every phrase on this list first — that's how the models learned them. In our scan of 30 top-starred, human-written SFW character cards, the phrases still turn up: "you feel a mixture of" appears in 4 of the 30, and the sample averages 0.07 lexicon hits per greeting. Slop judgment works on clusters: one "murmured" is prose, five murmurs and a knowing smile and a breath someone didn't know they were holding in the same scene is a fingerprint. Q: What is the most reliable AI tell in English roleplay prose? A: Not a word — a construction. The "not X, but Y" contrast pattern is the most-cited single marker: EQ-Bench weights contrast patterns as a standalone 25% of its slop score, and NousResearch's anti-slop guide calls it the number-one LLM rhetorical crutch. Behind it: emotion-cocktail formulas ("a mix of anticipation and dread") and fishing-line endings ("the choice is yours"). And a negative result worth knowing: em-dash density does not separate human from AI in English RP prose — we measured that separately. Q: Do slop word lists work as AI detectors? A: As lint, yes; as a lie detector, no. In our calibration, human top-card greetings averaged 0.07 hits while raw single-shot LLM greetings averaged 0.22 — the direction is right but the gap is small, because 2026 flagship models produce near-zero word-list hits in a single polished greeting. The list earns its keep in long multi-turn chats, on weaker models, and as a writing-discipline gate for card authors. Treat a clean scan as "not obviously sloppy," never as "human." ### How to Write AI Character Cards: A 2026 Guide Calibrated on 30 Top Cards URL: https://foreverse.app/blog/how-to-write-ai-character-cards Published: 2026-07-21 · Author: Deng Binjie We pulled the 30 top-starred SFW character cards from a major card hub and measured everything: first messages run a median of 178.5 words, 0 of 30 use HTML, 67% write the description in plain prose, 70% ship example dialogues, only 20% carry a lorebook. This guide turns those numbers — plus the community guide canon — into a writing procedure: format, token budget, the no-user-actions rule, greetings, example dialogs, and the anti-slop pass. Q: How long should a character card's first message be? A: 150–350 words is the working band for a dialogue card. Across the 30 top-starred SFW cards we sampled in July 2026, first messages ran a median of 178.5 words (P25=123, P75=219); a 40-word opener still charted in the top 15. The real rule is that the model copies the first message's length and style more than any other field, so write the greeting the length you want the replies to be. RPG and scenario cards can stretch to about 550 words; past 600 you're paying tokens to train the model to ramble. Q: How many tokens should a character card use? A: Keep permanent tokens — description, personality and scenario, which are re-sent with every message — between 800 and 1,500, and treat 2,000 as the ceiling. JanitorAI's two official tutorials cap at 2,000–2,500; the pixi guide recommends 500–1,000 for a single-character card; the top cards we measured had a full-card median of 1,085 tokens. Greetings and example dialogues don't count against this budget because they get pushed out as chat history grows. Q: What format should the description use — prose, PList, or W++? A: Plain prose. 67% of the top cards we sampled use it, platform docs recommend it, and no card in the sample uses strict PList, JSON, or interview format as its main definition. W++ survives only in 2023-era cards. Key-value lines keep one legitimate niche: enumerable info like appearance and clothing. Write personality, relationships and history as prose; list only the inventory-like facts. Q: Why shouldn't the card speak or act for the user? A: It's the single most-hated defect in English roleplay communities — JanitorAI's official help calls it users' most common and most irritating problem, and 25 of the 30 top cards observe the rule strictly. Every user action you write in the greeting teaches the model that taking over the user's character is allowed. Fix it by subject inversion: instead of "You are surprised to see him," write "The sight of him here doesn't fit." The one folk exemption is the RPG opener's passive hand-off ("You wake up in…") — scene only, never decisions or dialogue. Q: Do I need example dialogues and a lorebook? A: Example dialogues: yes for most cards — 70% of the top cards carry them, delimited with , and they are the most reliable way to lock a voice (heavy-accent cards run double-digit groups). A lorebook: only when the card has something to reveal conditionally — 20% of top cards carry one, and 5 of those 6 are RPG or scenario cards. A plain dialogue card with no secrets doesn't need one. ### Does a Bigger Model Fix AI Character Amnesia? A 400-Turn Companion Memory Test (DeepSeek v4 vs Qwen 3.8/3.7) URL: https://foreverse.app/blog/does-a-bigger-model-fix-amnesia-en Published: 2026-07-21 · Author: Deng Binjie “My AI roleplay partner forgets everything” is the most common complaint in companion chat, and the folk remedy is always the same: switch to a bigger model. We put that remedy through a 400-turn companion conversation with 50 planted memory probes, three models under one protocol: deepseek-v4-pro, qwen3.8-max-preview, qwen3.7-max. On bare context all three vendors are amnesiac (19–22%, barely distinguishable). Add retrieval-injected memory and they fan out into tiers — 89%, 74%, 59%. Ask about things never said, and even the most honest one passes only 5 of 7. The upgrade dividend is real — just not where users think it is. Tested 2026-07-21. Q: Which model has the best memory for AI roleplay? A: Of the three we tested, deepseek-v4-pro takes both crowns: 89% memory utilization (qwen3.8 74%, qwen3.7 59%) and 5-of-7 honesty on questions it had no answer for (the Qwen pair managed 2/7 and 0/7). State the boundaries too: one scenario (a 400-turn companion chat), one window size (16 turns), 27 clean probes — an engineering observation, not a universal leaderboard. And note that on bare context all three score 19–22%: without a memory system, every one of them is amnesiac. More models are queued for this series. Q: Context windows are at 1M tokens now — doesn't that mean the model can just remember everything? A: A window solves “fits,” not “gets used” — and you pay for it every turn. Chat resends history with each reply, billed per token: resending hundreds of turns wholesale is easily a hundred times the input of injecting a few retrieved entries. The repeatedly replicated “Lost in the Middle” research also shows models tend to overlook information buried mid-context. Real products run a sliding window; anything outside it only comes back through a memory system. Q: How does Foreverse deal with character amnesia? A: Two ways. Tavern-style chats get a memory plugin — a fact store plus retrieval that injects entries relevant to the current topic into the prompt, with an anti-fabrication instruction at the end of the injected block. Companions get a memory profile: every fact lives in a file on your phone, listed entry by entry — readable, editable, deletable, and exportable when you switch devices. The 19%-to-89% jump in this test is exactly the jump that design bets on. Q: How do I test whether my own app or model forgets? A: Plant hooks. Early in a chat, mention three small things with concrete values — a pet's name, an anniversary, a number. Dozens of turns later, ask about them in different words. Then ask about something you never said. Score two things only: did it get the planted facts right, and did it dare to admit not knowing the unplanted one. Most models fail the second test; treasure the ones that pass. ### 换个大模型，AI 角色就不失忆了吗？400 轮陪伴对话的记忆实测（DeepSeek v4 vs Qwen 3.8/3.7） URL: https://foreverse.app/zh/blog/does-a-bigger-model-fix-amnesia Published: 2026-07-21 · Author: Deng Binjie 「TA 又忘了我说过的话」是角色扮演和 AI 伴侣用户最痛的一刀，社区药方出奇一致：换个更强的模型。我们用 400 轮伴侣对话、50 道埋好的记忆考题，同协议考了三个模型：deepseek-v4-pro、qwen3.8-max-preview、qwen3.7-max。裸上下文三厂全失忆（19% 到 22%，几乎无差异）；加上检索记忆注入后分层拉开，89%、74%、59%；问从未提过的事，最诚实的一位也只过 7 道里的 5 道。升级的红利真实存在，只是不在用户以为的地方。测试日期 2026-07-21。 Q: 到底换哪个模型记性最好？ A: 本轮三个模型里，deepseek-v4-pro 双项第一：记忆利用率 89%（qwen3.8 74%、qwen3.7 59%），弃权诚实度 5/7（qwen 系 2/7 和 0/7）。边界也要说清：单场景（400 轮伴侣 IM）、单窗口档（16 轮）、27 道干净题，这是工程观察不是通用排行榜。且注意三家裸上下文都只有 19% 到 22%——不配记忆系统，换谁都失忆。更多模型的横评在排期中，跑完会更新到这条评测线里。 Q: 上下文窗口都到 1M 了，不是号称能记很多吗？ A: 窗口解决「装得下」，不解决「用得上」，而且付不起。聊天每轮都要把历史重发一遍、按 token 计费：几百轮历史全量重发，输入量是检索注入几条记忆的上百倍；「Lost in the Middle」这条被反复复现的研究还表明，长上下文中段的信息本来就容易被模型忽视。真实产品都是滑动窗口，窗口之外的事实只能靠记忆系统递回来。 Q: Foreverse 怎么解决角色失忆？ A: 两条线。酒馆聊天有记忆插件，事实记忆加检索双引擎，按当前话题把相关条目注入 prompt，注入块尾部带一行反编造兜底指令；AI 伴侣有记忆档案，TA 记了什么逐条列在你手机本机的文件里，看得到、改得了、删得掉，换手机整包带走。这轮实测里 19% 到 89% 的跳变，就是这套设计押注的那一跳。 Q: 我怎么测自己用的 App 或模型忘没忘？ A: 埋钩子。聊天早期自然地讲三件带具体值的小事（宠物名字、纪念日、一个数字），几十轮之后换个措辞回头问；再问一件你从未说过的事。判分只看两条：说过的答没答对，没说过的敢不敢承认不知道。第二条大多数模型都过不了，谁过了值得珍惜。 ### Qwen 3.8 vs Kimi K3 for Fiction: Two Trillion-Scale Models, Three Days Apart, One Blind Exam — 3:20 URL: https://foreverse.app/blog/qwen-3-8-vs-kimi-k3-fiction-en Published: 2026-07-21 · Author: Deng Binjie Two Chinese labs shipped trillion-scale models within three days: Kimi K3 (2.8T parameters) on July 16, Qwen3.8-Max-Preview (2.4T, self-described as second only to Fable 5, no independent benchmarks attached) on July 19. We ran the first paired blind fiction exam between them: same two novels, same anchor point, same instructions, 20 consecutive continuation rounds each, six AI judges under flipped mappings. Final score 3:20 — 1:10 on the fantasy epic, 2:10 on the palace novel, and every one of Qwen's three ballots contradicted itself when the mapping flipped. Both sides have documented flaws: Qwen3.8 was cited for a severe plot rewind and a modern-literary metaphor in period prose; K3 was cited for formulaic plot recycling, on top of its straight-quote and always-on-reasoning habits. Preview models are moving targets; every conclusion here is pinned to the hosted endpoint as of 2026-07-21. Q: K3 won 20 of 23 ballots — why would anyone still write fiction with Qwen 3.8? A: Ecosystem and marginal cost. Qwen3.8-Max-Preview is not sold per token; it ships inside Alibaba's subscription products (Token Plan, the Qoder platforms), so anyone already paying the subscription uses it at no extra charge, with preview-period credit consumption at 10% of standard rates and a further 80% off overnight on the Personal plan. It is also a genuine upgrade over its own predecessor: under the same protocol it beat Qwen 3.7 11:1 on the palace novel and 7:5 on the fantasy epic. But the head-to-head verdict stands — 3:20. Making it your primary continuation model is not a position the data supports. Q: How do Qwen 3.8 and Kimi K3 prices compare? A: Different billing shapes — don't force a per-token conversion. K3 is metered: $3.00 per million input tokens on cache miss, $0.30 on hit, $15.00 per million output, with always-on reasoning billed at full output price; our two 20-round chains cost ¥12.21 (about $1.70) total. Qwen3.8-Max-Preview is subscription-only: Token Plan's Lite tier lists at 39 CNY a month at the limited-time price (about $5.40), preview usage burns credits at 10% of standard rates, and our entire experiment did not exhaust the lowest tier's quota. One is à la carte, the other is a buffet — check your own usage shape first. Q: How do I use each model's strengths in Foreverse? A: Switch models per paragraph. Send the scenes you care about to K3 — its style fidelity holds 20 ballots — and accept the half-minute reasoning wait per round. Run daily progress and throwaway drafts on Qwen 3.8 if you already hold the subscription. When a stretch goes wrong, switch models and regenerate that one passage: plot rewinds and repetition loops are cumulative diseases, and one cut breaks the cycle. Both models connect with your own API key. Q: What happens when K3's weights go open on July 27? A: Moonshot has committed to releasing the full 2.8-trillion-parameter weights by July 27, 2026. Once they land, third-party hosts can compete on price and latency against the official API. Running it yourself is a different story — the official guidance starts at dozens of accelerators per node, so this is a hosting-market event, not a laptop event. Qwen 3.8's open-weight release has a date of "soon" and nothing firmer. When either model exits preview or the weights ship, we re-run the same protocol and update this page; the date at the top is the authority. ### Qwen 3.8 对上 Kimi K3：三天内连发的两个万亿级模型，谁更会写小说？40 轮双盲 3:20 URL: https://foreverse.app/zh/blog/qwen-3-8-vs-kimi-k3-fiction Published: 2026-07-21 · Author: Deng Binjie 中国两家厂商三天内连发万亿级模型：7 月 16 日 Kimi K3（2.8 万亿参数），7 月 19 日 Qwen3.8-Max-Preview（2.4 万亿参数，自称仅次于 Fable 5、未附独立 benchmark）。我们把两家放进同一张续写考卷：同两本书、同起笔点、同一套指令，各连续续写 20 轮，六评委双映射成对盲评。比分 3:20——玄幻 1:10、女频 2:10，qwen3.8 名下三张票还全部在映射翻转下自相矛盾。双方的坑都有票据：qwen3.8 被点名「严重剧情倒带」和「把夜色熬成了毒」，K3 被点名「舒痕胶的字据、夹袋」套路化复读，外加半角引号与思考关不掉两个存量病。preview 是移动目标，结论绑定 2026-07-21 的托管端点。 Q: Kimi K3 赢这么多，为什么还有人用 Qwen 3.8 写小说？ A: 场景和生态。qwen3.8-max-preview 目前不单独按量计费，随阿里自家产品走（Token Plan 订阅、Qoder 系平台），已经付了订阅的人用它不另花钱，preview 期 Credits 按 1 折计、个人版夜间再叠 2 折；它对自家前代也是实打实的升级：同协议双盲女频 11:1、玄幻 7:5，qwen3.7 的「恒定精修感官流」老毛病收敛大半。但同场双盲 3:20 也是真的：把它当续写主力，目前没有数据支撑。 Q: Qwen 3.8 和 Kimi K3 的价格怎么比？ A: 计费形态不同，别硬折算单价。K3 按量计费：官方牌价每百万 token 输入未命中缓存 ¥20、命中 ¥2、输出 ¥100，思考 token 照输出价收，我们实测两本书各 20 轮共花 ¥12.21。qwen3.8-max-preview 不单卖，走 Token Plan 订阅（个人版限时 39 元/月起），preview 期 Credits 消耗按 1 折计、每晚 22 点到次日 8 点再叠 2 折，我们整轮实验没耗尽最低档订阅额度。一个按杯卖、一个自助餐，先看自己的用量形态再选。 Q: 在 Foreverse 里怎么让两个模型各用各的长处？ A: 续写逐段可换模型。关键剧情交 K3，它的文风保真有 20 张票背书，代价是每轮半分钟的思考等待；日常推进、支线试写用订阅额度里的 qwen3.8 跑量。哪段写崩了就地换模型重写一段——剧情倒带和复读都是累积病，切一刀就断。两家都支持自带 key 接入，同一本书里混用是现成操作。 Q: 7 月 27 日 K3 开源权重之后会怎样？ A: 月之暗面承诺 7 月 27 日前放出 2.8 万亿参数的全量权重。落地后第三方推理商可以竞价托管，价格和延迟大概率比官方 API 松动；但这个量级个人本地跑不现实，官方口径的推理配置就要几十张加速卡起步。qwen3.8 的开源只说了「soon」，没给日期。权重落地或 preview 转正，任一发生我们都按同协议复测并更新本页，以页首日期为准。 ### Is Qwen 3.8 Good at Writing Fiction? 40 Blind Rounds Against Qwen 3.7, 48 Hours After Launch URL: https://foreverse.app/blog/qwen-3-8-fiction-20-rounds-en Published: 2026-07-21 · Author: Deng Binjie Qwen3.8-Max-Preview shipped July 19, 2026; within 48 hours we ran it through the same protocol as our Kimi K3 test — two Chinese novels, 20 consecutive continuation rounds each, paired double-blind against its predecessor Qwen 3.7 Max. The palace novel flipped 11-1 for the new model; the fantasy epic barely moved at 7-5 with a 6-6 opening window. The twist: structural statistics say 3.7 is closer to the originals, and 3.8 writes the most uniform sentence lengths we have ever measured — yet the blind vote still went to 3.8. Honest caveats: sensory pile-ups in place of plot, a 20-round fantasy deadlock, one 'brewing the night into poison' metaphor in a period novel, and always-on reasoning eating 50-90% of output tokens. All results pinned to the 2026-07-21 hosted preview. Q: Is Qwen 3.8 better than Qwen 3.7 for fiction? A: Depends on the genre. On the palace-intrigue novel it is a rout: same protocol, 20 consecutive continuation rounds each, six heterogeneous AI judges under two flipped blind mappings, 11-1 overall with every review window at 11-1 or 11-0 — a clear break from 3.7's eighth-of-nine placement on that book. On the fantasy epic it is a squeak: 7-5 overall, with the opening window tied 6-6. The predecessor's signature polished sensory stream has receded, but judges still flagged sensory pile-ups and a 20-round plot deadlock. All results are pinned to the hosted preview as of July 21, 2026. Q: Is Qwen 3.8 better than Kimi K3 for fiction? A: No. In paired double-blind under the same protocol, Qwen 3.8 lost to Kimi K3 3-20 across the two books combined (1-10 on the fantasy epic, 2-10 on the palace novel), and K3 remains the current continuation champion. Full ballots, judge comments, and the cost comparison live in our separate Qwen 3.8 vs Kimi K3 head-to-head; this page is about the vertical 3.7-to-3.8 comparison. Q: How much does Qwen 3.8 cost, and how fast is it? A: We ran it on an Alibaba Token Plan subscription (verified 2026-07-21): the personal Lite tier is 39 CNY/month at the limited-time price, and qwen3.8-max-preview credits are metered at 10% of standard burn during the preview promotion. Measured speed: 12 to 38 seconds per round, reasoning always on, thinking tokens taking 50-90% of completion output. The whole experiment — four chains, 80 rounds, 1,227,401 input tokens and 103,229 output tokens, zero cache hits — did not exhaust the Lite quota. Q: Will these results hold for the final Qwen 3.8 release? A: No promises. Alibaba's documentation states the preview model keeps iterating and will be taken offline or replaced by a production version afterward, so every vote and metric on this page is pinned to the hosted preview as of July 21, 2026. An open-weight final release has been announced without a date; when it ships, we will re-run the same 40 blind rounds, fold the result into our evergreen model guide, and pin a pointer to the new test at the top of this page. ### Qwen 3.8 写小说怎么样？发布 48 小时的 40 轮双盲实测：女频 11:1 碾压前代，玄幻只赢一口气 URL: https://foreverse.app/zh/blog/qwen-3-8-fiction-20-rounds Published: 2026-07-21 · Author: Deng Binjie Qwen3.8-Max-Preview 发布 48 小时的增量实测：沿用 Kimi K3 那轮的同一套协议（同两本书、同起笔点、同上下文说明），与前代 Qwen 3.7 Max 各连续续写 20 轮成对双盲。票数女频 11:1 碾压、玄幻 7:5 险胜，前代的「精修感官流」在古言场明显收敛。最有意思的反差：结构统计反而是 3.7 更贴原著，句长 CV 全场最平是 3.8 的新体征，盲评却照样判 3.8 赢。三个坑如实记录：感官堆砌替代叙事推进、玄幻 20 轮卡进死循环、古言里写出「把夜色熬成了毒」。preview 是移动目标，全部结论绑定 2026-07-21 的端点。 Q: Qwen 3.8 写小说比 3.7 强多少？ A: 分文体。宫斗古言场是碾压：同协议各连续续写 20 轮，六个异构 AI 评委、双映射成对盲评，整体票 11:1，三个评审窗口全部 11:1 或 11:0，前代在九模型榜上垫底区的第八名被明显甩开。传统玄幻场只是险胜：整体票 7:5，开局窗口还打成 6:6 平手。前代标志性的「精修感官流」收敛了，但感官偏科与剧情死循环仍被评委点名。全部结论绑定 2026-07-21 的 preview 端点。 Q: Qwen 3.8 和 Kimi K3 写小说谁更强？ A: 同一套协议的成对双盲里，Qwen 3.8 对 Kimi K3 两本书合计 3:20 惨败（玄幻 1:10、女频 2:10），K3 仍是当前续写强者。完整票据、评语与成本对照在我们的《Qwen 3.8 对 Kimi K3》对决篇里单独展开，本文主轴是 3.7 到 3.8 的纵向对比。 Q: Qwen 3.8 的 API 多少钱？一段续写要等多久？ A: 我们跑在阿里 Token Plan 订阅上（2026-07-21 核实）：个人版 Lite 限时 39 元/月，qwen3.8-max-preview 预览期 Credits 按 1 折消耗。速度实测单轮 12 到 38 秒，思考关不掉，reasoning token 占输出的 50% 到 90%。整场实验四条链 80 轮共 1,227,401 个 input token、103,229 个 output token，缓存命中为零，没有耗尽 Lite 当期额度。 Q: 正式版发布后，这份结论还算数吗？ A: 不承诺。官方文档写明 preview 期间模型会持续迭代，预览结束后下线或替换成正式版本，所以本页票数与指标只绑定 2026-07-21 的 hosted preview。官方已预告正式版将开源；等它转正，我们会复跑同一套 40 轮双盲，把名次更新进长青选型页，并在本页顶部挂新实测的导流条。 ### Your SillyTavern Backup Zip, Unpacked: What Actually Moves to the Phone URL: https://foreverse.app/blog/sillytavern-backup-to-phone Published: 2026-07-19 · Author: Deng Binjie Desktop SillyTavern's Download Backup produces one zip of your whole tavern. An item-by-item audit of what Foreverse restores from it — characters, chats as archived branches, worldbooks, presets, Quick Replies, regex — and what deliberately stays behind. Q: Where is the backup button in desktop SillyTavern? A: User Settings → Download backup. It zips your entire data// directory — characters, chats, worldbooks, presets, Quick Replies, extension settings — into one file. That zip is the input for the whole migration. Q: Do my chat histories survive the move? A: Yes — that is the point of using the backup zip instead of re-importing cards one by one. Each character's chats/*.jsonl files are restored as archived chat branches on that character, so a 400-message history from the desktop is scrollable on the phone. Q: My zip has an extra folder layer inside. Will it still import? A: Yes. The importer expects ST's own layout (characters/, chats/, worlds/ at the zip root) but also detects the common case where you zipped the parent folder by hand, and strips the shared prefix automatically. Q: Is anything read that shouldn't be? A: secrets.json — the file where desktop ST keeps your API keys — is never read. Keys don't belong in a migration; on the phone you re-enter them once and they live AES-encrypted in the device Keystore. Q: Is zip import free? A: Full-backup zip import and export sit in the Pro tier (single cards, links, and folder scanning are free within the three-card limit). One backup restore on moving day is the typical use. ### How to Import Character Cards on Android: Five Ways In, From One PNG to a Whole Backup URL: https://foreverse.app/blog/import-character-cards-android Published: 2026-07-19 · Author: Deng Binjie A field guide to getting character cards onto your phone: opening a PNG or .charx from your file manager, folder scanning, pasting a chub or JanitorAI link, the community hub, and the full SillyTavern backup route — plus the three failure modes that account for most broken imports. Q: Do cards from desktop SillyTavern work unchanged? A: Yes. Cards are standard chara_card_v1/v2/v3 files, not a private format. A PNG or JSON exported from desktop SillyTavern imports on the phone with greetings, embedded lorebook, and example dialogue intact — and cards you edit on the phone export back out the same way. Q: Why does my PNG import as a blank character? A: Because something re-encoded the image on the way to you. Card data lives in PNG tEXt metadata chunks; messenger compression and photo-library optimization rebuild the pixels and silently drop those chunks. The portrait survives, the character doesn't. Ask the creator for the JSON version, or download from the original card page instead of a chat attachment. Q: What is a .charx file? A: The chara_card_v3 spec's zip container: card.json plus bundled assets like emotion sprites or audio. Foreverse parses it the same way desktop SillyTavern 1.18 does. If a file has no extension but is actually a zip, the importer peeks inside for card.json before deciding — so mislabeled charx files still land. Q: How many cards can I import for free? A: Three character cards on the free tier, from any source — files, links, or the community hub. Beyond that, card slots are part of the Pro tier. Imported cards stay yours either way: they are standard-format files you can export in full at any time. Q: Can I bring chat history too? A: Not through a single card file — a card carries the character, not your conversations. Chat history comes across inside a full SillyTavern backup zip, which restores each character's chats as archived branches. That route is covered in its own guide. ### 酒馆备份 zip 拆包记：从电脑搬到手机，哪些东西真的跟着走 URL: https://foreverse.app/zh/blog/st-backup-zip-migration Published: 2026-07-19 · Author: Deng Binjie 桌面 SillyTavern 的 Download backup 会把整个酒馆打成一个 zip。逐项拆开看 Foreverse 从里面恢复什么：角色卡、聊天记录变归档分支、世界书、预设、Quick Reply、正则脚本，以及哪两样东西是故意不搬的。 Q: 桌面酒馆的备份按钮在哪？ A: User Settings → Download backup。它把 data// 整个目录打成一个 zip，角色、聊天、世界书、预设、Quick Reply、扩展设置全在里面。迁移的输入就是这一个文件。 Q: 聊天记录能搬过来吗？ A: 能，这正是用备份 zip 而不用逐张重导卡的理由。每个角色的 chats/*.jsonl 会恢复成该角色名下的归档聊天分支，桌面上攒了几百楼的历史在手机上照样能翻、能接着聊。 Q: 我打包时手滑多套了一层文件夹，还能导吗？ A: 能。导入器按酒馆原生结构（characters/、chats/、worlds/ 在 zip 根）解析，也自动识别「把上级文件夹整个压进去」的常见形态，公共前缀会被剥掉。 Q: API Key 会跟着 zip 走吗？ A: 不会。存密钥的 secrets.json 从头到尾不被读取。Key 在手机上重新填一次，AES 加密存进系统 Keystore。会悄悄搬凭据的迁移工具，才是需要警惕的迁移工具。 Q: zip 整包导入免费吗？ A: 整包备份导入导出属于 Pro 权益。单张卡、链接导入、文件夹扫描在三张免费额度内都免费；搬家日用一次整包恢复是它的典型场景。 ### Kimi K2.6 要不要升级 K3？先看 21 比 1 的双盲，再算 3.7 倍的价差 URL: https://foreverse.app/zh/blog/kimi-k2-6-to-k3-upgrade Published: 2026-07-18 · Author: Deng Binjie K3 发布后，开放平台首页挂着最高返券 30% 的充值活动，模型列表里 moonshot-v1 和 kimi-k2.5 标了 8 月 31 日全平台下线，夹在中间用 K2.6 写书的人最纠结。我们刚跑完 K3 对 K2.6 的同协议成对双盲：两本书 24 票，K3 拿 21 票、K2.6 拿 1 票，连 K2.6 引以为傲的后程都没守住。这篇按三类用户拆：长跑党看双盲实证，预算党算输入 3 倍输出 3.7 倍的差价加思考税，观望党先确认 K2.6 根本不在下线清单里。 Q: K2.6 和 K3 写小说差距大吗？ A: 同协议成对双盲的答案是大。同两本书（玄幻加宫斗古言）、同起点各连写 20 轮，6 个异构评委、每书两套匿名映射共 24 票：K3 拿 21 票，K2.6 拿 1 票，2 票无效，且唯一反对票在映射翻转下自相矛盾，属位置偏差。按 early、mid、late 三窗逐窗计票，玄幻场三窗全是 10 比 0、女频场三窗全是 11 比 1，K2.6「越写越收敛」的后程优势没有在任何一个窗口扳回局面。 Q: K2.6 会像 moonshot-v1 一样下线吗？ A: 官方现状：不在清单里。8 月 31 日全平台下线的是 moonshot-v1 全系和 kimi-k2.5（已停止向新注册用户开放），K2.6 仍列在官方模型列表的在售主力位，定价页照常挂牌。官方没有公布过 K2.6 的退役时间表，我们不替它猜。可以确定的是 Moonshot 清场有节奏：kimi-k2 系列 5 月 25 日下线、v1 系 8 月底下线，用旧模型名的代码每季度值得全局搜一遍。 Q: 从 K2.6 升到 K3 要改什么？ A: 代码层面一行：model 字段从 kimi-k2.6 改成 kimi-k3，base_url 和 key 都不动。预算和习惯层面有三件事：K3 思考恒开且当前只有 max 一档，关不掉，每轮多等几十秒；思考 token 按输出价 100 元每百万计费，我们实测输出的 81% 到 86% 是思考过程；输出预算要给思考留量，别按可见正文的字数设上限。 Q: 充值返券活动值得等吗？ A: 活动期 2026 年 7 月 15 日到 8 月 11 日，单笔充值满 99 元按阶梯返代金券，5000 元及以上档最高返 30%，赠券 90 天有效，具体档位以开放平台充值页说明为准。反正要充的，8 月 11 日前充合算；不建议为凑返券档位囤钱，Kimi 开放平台充值不支持退款，赠券又有有效期，按你 90 天内真实写得完的量充。 ### Kimi K3 和 DeepSeek 哪个好？写小说的人先看这份对表，再等对决 URL: https://foreverse.app/zh/blog/kimi-k3-vs-deepseek-v4-for-fiction Published: 2026-07-18 · Author: Deng Binjie K3 发布 48 小时，对比文全在刷代码榜，写小说角度没人给数据。我们把 K3 刚跑完的两书 20 轮续写链和 DeepSeek 双档的同书归档放上同一把尺：玄幻场 K3 综合贴近度 0.388，落在 V4 Flash 的 0.356 与 V4 Pro 的 0.454 之间；女频场 0.396，未及 V4 Pro 的 0.280；成本这局 DeepSeek 碾压，一段 K3 的钱够 Flash 谷时写二十多段。必须说明：这是同尺隔空对表，不是同场盲评，成对对决在排期中。 Q: Kimi K3 和 DeepSeek 哪个写小说更好？ A: 目前只能给对表，给不了胜负。同一把尺（同两本书、同起点、同 20 轮连续续写、同一套指标定义）下：玄幻场 K3 综合贴近度 0.388，介于 DeepSeek V4 Flash（0.356，该场盲评冠军）与 V4 Pro（0.454）之间；女频场 K3 0.396，未及 V4 Pro 的 0.280。但两家从未进过同一场盲评，谁击败谁没有考卷。按预算和场景选：日常长跑选 DeepSeek 按文体双档，想要 K3 的纹理复刻力再为它付溢价。 Q: 为什么不直接盲评分个胜负？ A: 成对盲评要求两份产物同场同协议、评委不知来源。K3 这轮的成对对手是同厂上代 K2.6（战绩 21 比 1），DeepSeek 的名次来自另一场九模型双盲，两场评审形态不同，跨榜硬比名次会混入协议变量。结构指标是同一套定义在同两本书上算出来的，所以可以对表；胜负要等 K3 与 DeepSeek 双档进同一场成对盲评，这场在排期中，跑完更新本页。 Q: K3 比 DeepSeek 贵多少？ A: 对 V4 Flash 谷时牌价，输入贵 20 倍、输出贵 50 倍；按实测每段成本折约 25 倍。K3 每百万 token 输入 20 元、输出 100 元，思考恒开且思考 token 照价计费，我们实测两本书各 20 轮共花 12.21 元，摊到每段约 3 毛，每轮等 37 到 54 秒；DeepSeek V4 Flash 谷时输入 1 元、输出 2 元，同样一段一分二厘左右，缓存命中价是未命中的 1/50。一段 K3 的钱在 Flash 谷时能写二十多段。 Q: K3 写中文小说有什么已知的坑？ A: 实测记录里有三个：两本书约 85% 的对白用半角引号（GLM 5.2 同款顽疾，中文网文读者一眼出戏）；写第一人称古言时「我」的密度是原著的 1.8 倍（每千字 16.2 次对 9 次）；思考模式恒开关不掉，等待和费用都躲不开。另有一处专名细节：它把皇后居所写成了剧版的「景仁宫」，小说设定是凤仪宫，说明底层记忆微偏电视剧语料。 ### Kimi K3 Creative Writing Test: 40 Rounds of Novel Continuation, 48 Hours After Launch URL: https://foreverse.app/blog/kimi-k3-fiction-20-rounds Published: 2026-07-18 · Author: Deng Binjie Kimi K3 shipped July 16, 2026; within 48 hours we ran it through the same protocol as our nine-model benchmark — two Chinese novels, 20 consecutive continuation rounds each, paired double-blind against its predecessor K2.6. The vote: 10-0 on the fantasy epic, 11-1 on the palace novel, and the single dissent contradicted itself under mapping flips. K2.6's metaphor pile-ups and verbatim self-copying did not recur. Three honest caveats: ~85% of dialogue in Western straight quotes (a disease K2.6 never had), first-person density 1.8× the original author's, and always-on reasoning that makes every round take 37-54 seconds. Total bill for the whole experiment: ¥12.21. Q: Is Kimi K3 good for fiction? A: Clearly better than its predecessor, with receipts: under the same protocol, K3 and K2.6 each continued two Chinese novels for 20 consecutive rounds, and six heterogeneous AI judges under two flipped blind mappings voted 10-0 (fantasy) and 11-1 (palace intrigue) for K3 — the single dissent contradicted itself when the mapping flipped. Sentence-length variance moved closer to the original author on both books, and K2.6's signature metaphor pile-ups and self-copying did not recur. Three caveats: ~85% of dialogue in Western straight quotes, first-person density 1.8× the original's on the palace novel, and 37-54 seconds per round from always-on reasoning. No head-to-head against DeepSeek yet, so no field-wide rank. Q: Should I upgrade from Kimi K2.6 to K3 for creative writing? A: The quality direction is unambiguous — 21:1 in paired double-blind — and so is the price of admission. International pricing moves from $0.95 to $3.00 per million input tokens and from $4.00 to $15.00 per million output; our measured cost per continuation went from under two cents to roughly four cents, every round waits half a minute longer, and K3 picked up a straight-quote habit K2.6 never had. If budget or latency matters, staying on K2.6 is defensible. If you are unsure, run the blind test we always recommend: same opening, three continuations from each model, read them shuffled, then flip the cards. Q: Does Kimi K3's 1M context window help with fiction? A: This entire 21:1 result happened inside a 16k-token window; the million-token window never entered the exam. Our five-tier control experiment showed that cramming a whole novel into context with no style instruction still produces boilerplate similes at ten times the original author's density — volume does not transfer style. Filling K3's window once costs about $3.15 of input before it writes a word. The big window earns its keep on one-shot whole-book tasks like summarization and character audits, not on repeated continuation. Q: How much does Kimi K3 cost per continuation? A: Official international pricing (verified 2026-07-18): $3.00 per million input tokens on cache miss, $0.30 on hit, $15.00 per million output, flat across the 1M window, currently max-reasoning only. Measured: our two 20-round chains, 41 requests including a connectivity probe and retries, cost ¥12.21 total (about $1.70) — roughly four cents per continuation. Note that 81-86% of output tokens were invisible reasoning, billed at full output price; visible prose ran about 250 Chinese characters per round. Q: Is Kimi K3 better than DeepSeek for fiction? A: No exam paper yet, so no verdict. Indirect evidence: on composite structural distance to the original author's profile (lower is closer), K3 scored 0.388 on the fantasy epic versus champion DeepSeek V4 Flash's 0.356, and 0.396 on the palace novel versus champion V4 Pro's 0.280 — inside the top tier's neighborhood on both books, ahead of the champion on neither. This round's judging format also differs from the nine-model leaderboard, so cross-table conversion stops there. A paired K3-versus-DeepSeek blind run is our next exam; until it lands, claims in either direction have no paper behind them. ### Kimi K3 写小说怎么样？发布 48 小时的 40 轮双盲实测：21:1 碾压前代，但有三个如实的坑 URL: https://foreverse.app/zh/blog/kimi-k3-fiction-20-rounds Published: 2026-07-18 · Author: Deng Binjie Kimi K3 发布 48 小时的增量实测：沿用九模型横评同一套协议（同两本书、同起笔点、同上下文说明），与前代 K2.6 各连续续写 20 轮成对双盲，票数玄幻 10:0、女频 11:1，唯一反对票在映射翻转下自相矛盾；K2.6 的比喻堆砌与整段自我复制没有复发，句长变化幅度两本书都更贴原著。三个坑如实记录：约 85% 对白用半角引号（K2.6 没有的新病）、女频「我」密度是原著近 1.8 倍、思考关不掉导致每轮 37-54 秒且八成输出是推理 token。K3 侧 41 次请求实付 ¥12.21。 Q: Kimi K3 写小说怎么样？ A: 显著强于前代 K2.6，有票可查：同协议下两本书各连续续写 20 轮，六个异构 AI 评委双映射成对盲评，玄幻场 10:0、女频场 11:1，唯一反对票在映射翻转下自相矛盾。句长变化幅度两本书都更贴原著，K2.6 标志性的比喻堆砌和整段自我复制没有复发。坑也有三个：约 85% 对白用半角引号、女频第一人称密度是原著近 1.8 倍、思考模式关不掉导致每轮要等 37 到 54 秒。与 DeepSeek 双档的成对对决还没跑，全场名次先不排。 Q: 写小说该从 K2.6 升级到 K3 吗？ A: 质量方向明确，代价也明确。文字上 K3 成对双盲 21:1 压过 K2.6，还治好了比喻堆砌；代价是输入单价从 ¥6.5 涨到 ¥20、输出从 ¥27 涨到 ¥100（每百万 token），实测一段续写成本从不到一毛涨到三毛上下，每轮多等半分钟，另染上 K2.6 没有的半角引号病。预算敏感或讨厌等待，留在 K2.6 不丢人；哪类用户该升、哪类不必，我们在「K2.6 要不要升级 K3」一篇里按三类用户分开算了账。拿不准就别赌：同一段开头新旧各续三段，盲着读完再翻牌。 Q: K3 的 1M 上下文窗口对续写小说有用吗？ A: 这场 21:1 的胜利全程发生在 1.6 万 token 窗口里，与 1M 无关。我们的五档对照实验证明过：把整本书塞进窗口、不加文风指令，产出照样满篇套话，密度是原著的 10 倍，文风不会自动变像；而装满一次 K3 的 1M 窗口光输入就约 ¥21。1M 真正的主场是整卷总结、全书人物梳理这类一口气读完全书的一次性任务，反复续写要的是按相关性挑出的一小撮上下文。 Q: Kimi K3 的 API 什么价，实测一段续写多少钱？ A: 官方牌价（2026-07-18 核实）：输入未命中缓存 ¥20、命中 ¥2、输出 ¥100，单位每百万 token，1M 窗口全程平价不分段，当前只有满档思考一个档位。实测：两条 20 轮链共 41 次请求实付 ¥12.21，平均一段约三毛；其中输出 token 的 81% 到 86% 是看不见的推理过程，可见正文每轮约 250 字，推理部分照常按输出价计费。 Q: K3 和 DeepSeek 谁更适合续写小说？ A: 还没有成对考卷，不硬答。间接证据：与原著结构画像的综合距离上，玄幻场 K3 0.388 对冠军 DeepSeek V4 Flash 0.356，女频场 K3 0.396 对冠军 V4 Pro 0.280，K3 两场都进了第一梯队的区间、两场都没有越过冠军。本轮评委形态与九模型榜不同，跨榜换算只能到这一步。K3 对 DeepSeek 的成对双盲是我们下一份考卷；跑完之前，「K3 吊打 DeepSeek」和「K3 不如 DeepSeek」都属于没有考卷的说法。 ### GLM-5.2 写小说怎么样？双文体唯一双前三的完整档案（40 轮双盲实测） URL: https://foreverse.app/zh/blog/glm-5-2-fiction-file Published: 2026-07-18 · Author: Deng Binjie 智谱 GLM-5.2 在我们两场双盲横评（890 万字传统玄幻 + 《甄嬛传》宫斗古言）里是唯一双榜都进前三的模型：玄幻场第 3、宫斗场 2-3，评语「限知观察质感最纯」。这页是它的完整成绩档案：两场名次与盲评评语原文、半角引号顽疾的跨书证据链、检测正则被它骗过的笑话，以及 GLM-5.5 传闻的如实口径——发布当天我们会复测更新本页。 Q: GLM-5.2 写小说怎么样？ A: 有双盲实测名次可查：在我们两场横评（890 万字传统玄幻 + 《甄嬛传》宫斗古言，各连续续写 20 轮）里，它玄幻场第 3、宫斗场 2-3，是九个模型里唯一双榜都进前三的。玄幻场评语「内容层贴近（战术商议、单字传讯都对味）」，宫斗场评语「限知观察质感最纯」。代价是跨文体复现的半角引号顽疾，两本书都没改掉。 Q: 智谱的模型适合写哪类文？ A: 数据支持的答案是「都不偏科，但都不是第一」。单一文体的峰值它拿不到：玄幻第一是 DeepSeek V4 Flash，宫斗第一梯队是 DeepSeek V4 Pro 和 GPT-5.6 Terra。它的独特位置是跨文体稳定：文体还没定型的书、文戏武戏都要的长跑，双前三是保险牌。宫斗场那条「限知观察质感最纯」的评语，也让它在第一人称限知视角的书里格外值得一试。 Q: GLM 写中文小说的半角引号问题能修好吗？ A: 我们没有证据支持「一句提示词根治」。这个习惯在玄幻场 20 轮里出现 108 处，宫斗场原样复发，两场盲评都按表层观感扣了分。可行的是绕行：提示词里显式加引号约束（值得一试但不打包票），或者对话密集的章节逐段换模型、对已生成段落编辑重生。哪天新版本修掉了，本页会随复测更新。 Q: GLM-5.5 什么时候发布？会更强吗？ A: 官方没有任何官宣。目前唯一的时间口径来自摩根大通 2026 年 6 月下旬的研报预测：8 月发布，参数量可能突破万亿。这属于投行预测而非事实，8 月也完全可能跳票。我们的承诺写在页面里：GLM-5.5 发布当天，用同一套双盲协议复测，更新本页的每一个名次。 ### Is a 1M-Token Context Window Worth It for Fiction? We Ran the Control Experiment URL: https://foreverse.app/blog/is-1m-context-worth-it-for-fiction Published: 2026-07-18 · Author: Deng Binjie Kimi K3, DeepSeek V4, Claude Fable 5 and GPT-5.6 all carry million-token windows as of July 2026 — a whole 500k-word serial fits in one request. Our five-tier control experiment says fitting is not learning: a model handed an entire novel with no instruction wrote clichés at ten times the original author's density, and one explicit sentence did what 200,000 tokens of raw prose couldn't. Plus the arithmetic: filling K3's window costs about $3.15 per press. Q: Is a 1M context window worth paying for if I read and continue fiction? A: Mostly no, if the plan is cramming the whole book in. The big windows solved "doesn't fit" — a 500k-word serial runs roughly 625k to 830k tokens and now fits with room to spare. They didn't solve "fits but gets skimmed" (OpenAI's own eval reports 73.8% retrieval at 512K-1M depth), and they didn't change what our control experiment showed: raw prose volume doesn't transfer style. Our 360-round continuation benchmark ran entirely inside 16k-token windows and never needed more. Q: Does a bigger context window make the AI write more like the original author? A: Not by itself. We ran a five-tier control from 4k to 200k tokens of original prose in context. With no style instruction, the model handed an entire 270k-character novel still wrote boilerplate similes at ten times the original author's density, in its own default voice. With one explicit baseline sentence — anchor style to the original passages — the direction held at 200k just as it did at 4k. The instruction did the work; the extra 196k tokens did not. Q: How much does it cost to actually fill a 1M window? A: Kimi K3 prices input at $3.00 per million tokens, flat across its 1,048,576-token window, so one cold full-window request costs about $3.15 before output. A 700k-token serial costs about $2.10 of input per continuation press; twenty presses in an evening is roughly $42, paying for the same unchanged chapters twenty times. Prefix caching drops repeat input to $0.30 per million — real relief, but only while the prefix stays byte-identical, and editing or regenerating a passage upstream resets it. Q: If not the whole book, what should the AI see? A: A structured slice: the always-on facts (protagonist, world rules) re-sent every turn, plus lorebook entries keyed to trigger words so each faction or subplot enters the request only when the current scene mentions it. That is how our own benchmark runs stayed inside 16k tokens across 360 rounds. The window arithmetic explains why cramming loses; relevance-gated feeding is what replaces it. ### Grok 4.5 for Fiction: We Already Had Its File — Two Novels, 40 Rounds, Blind-Judged URL: https://foreverse.app/blog/grok-4-5-fiction-test-file Published: 2026-07-18 · Author: Deng Binjie Every Grok 4.5 review benchmarks coding. Nobody answers the question people actually type into search: is it any good for stories? As it happens, Grok 4.5 was one of nine models in our long-run continuation benchmark — 20 consecutive rounds on each of two novels, ranked by double-blind review. The file shows a twice-reproduced repetition loop, a dead-last palace-intrigue placement from both reviewers, a first-person density at half the original's — and the scenarios where it is genuinely fine. Q: Is Grok 4.5 good for creative writing? A: In short bursts, yes — brainstorms, dialogue rewrites, modern colloquial registers. Over long runs, no: in our 20-round continuation benchmark it locked into verbatim repetition loops in both books we tested. On the palace-intrigue novel, both blind reviewers ranked it last of nine, with twenty rounds of story time never leaving one afternoon. The failure is cumulative, so a quick demo will look fine. Q: Should I use Grok for roleplay? A: Long roleplay is exactly the regime where its documented failure lives: dozens of consecutive turns, each output feeding back into context. In our data Grok 4.5 hit repetition loops in both a fantasy epic and a palace-intrigue novel under that setup. If you run it anyway, keep sessions short, watch for sentences repeating verbatim, and switch models the moment the story clock stops advancing. Short in-character exchanges in modern registers are a much safer fit than long ornate-period campaigns. Q: Why does Grok keep repeating itself when writing long stories? A: It's the classic autoregressive fixed point: as the context fills with the model's own text, imitating itself becomes easier than advancing the plot. Most models resist it better — under the same prompt and protocol, Grok 4.5 was the only system in our field that froze in both test genres. One fantasy stretch repeats the same two paragraphs three times word for word. Prompt tweaks don't cure it; switching models or regenerating the looping segment breaks the cycle. Q: Did you test the released Grok 4.5 or an older checkpoint? A: The released model. Broad release was July 8, 2026; the runs behind our published benchmark posts landed in mid-July 2026, within days of it. Placements are bound to the version we called — xAI iterates, and a future update could change the picture. We re-run the benchmark as models change, and the date at the top of this page is the truth source. ### Claude Fable 5 写小说值不值？价格账、英文圈盲测证据和一个我们还没补上的缺口 URL: https://foreverse.app/zh/blog/claude-fable-5-for-fiction-worth-it Published: 2026-07-18 · Author: Deng Binjie Claude Fable 5 是当前最贵的通用可用模型：每百万 token 输入 $10、输出 $50，7 月 20 日起 Claude 订阅也不再内含，只能按量买 credits。英文写作圈的盲测把它叫最强 raw writer，但它 6 月停服过 19 天，而我们的九模型中文续写横评里没有它。这篇把价格实算、证据分层和决策树摊开：什么人现在就值、什么人该等实测、什么人用 DeepSeek 就够。 Q: Fable 5 怎么收费，多少钱？ A: API 牌价每百万 token 输入 $10、输出 $50，缓存命中 $1（输入价的十分之一）。2026 年 7 月 20 日起 Claude 订阅（Pro/Max/Team）不再内含 Fable 5，订阅用户要继续用需在设置里开通按量 usage credits，费率与 API 相同，另有每日 $2,000 的兑换上限。它是 Opus 4.8（$5/$25）的两倍、Sonnet 5 介绍价（$2/$10）的五倍，是 Anthropic 现行价目表上最贵的通用可用模型。 Q: Fable 5 写中文小说强吗？ A: 诚实答案：没有可引用的数据。英文侧证据不少（Noren 的盲测把它评为最强 raw writer，对 GPT-5.6 Sol 的 64 产物盲评八组全胜），但这些全是英文写作场景。我们的九模型中文续写横评（累计 360 轮、双盲评审）没有包含它，中文长跑表现目前是空白。英文证据不能平移到中文：同一个模型换个文体名次都会大挪移，换语言更不该默认。我们的 20 轮续写链已排上，测完会更新本文。 Q: 和 DeepSeek 比贵多少？ A: 按一次典型续写算（输入 1 万 token 未命中缓存、输出 800 token）：Fable 5 约 $0.14，DeepSeek V4 Flash 约 $0.0016，差 86 倍。写完 20 万字（约 500 段），Fable 5 约 $70（¥500 上下），V4 Flash 约 ¥6。单价层面输入差 71 倍、输出差 179 倍。这不构成「别用 Fable 5」的结论，但它决定了合理用法：关键章节单点召唤，其余段落交给便宜档。 Q: 为什么你们的横评没测 Fable 5？ A: 时间线错开了：它 6 月 9 日才发布，6 月 12 日起因出口管制全球停服 19 天，7 月 1 日恢复时我们两场横评的模型清单和生成链已经定型。这是榜单里最显眼的缺席，我们不打算用英文证据糊弄过去——同协议的 20 轮中文续写链已排上日程，结果出来第一时间补进横评答案页和本文。 ### AI Model Sunset Calendar, Summer 2026: deepseek-chat Retires July 24, Moonshot V1 Ends August 31 — What to Switch To URL: https://foreverse.app/blog/api-sunset-calendar-summer-2026 Published: 2026-07-18 · Author: Deng Binjie At least six model shutdowns and price changes land between July 20 and August 31, 2026: Claude Fable 5 moves to usage credits on July 20, deepseek-chat and deepseek-reasoner stop resolving on July 24, GitHub Models retires entirely on July 30, Moonshot V1 and kimi-k2.5 sunset on August 31, and Sonnet 5's introductory pricing ends the same day. Every date verified against official announcements, with replacement picks for fiction-continuation users backed by our 360-round benchmark. Q: deepseek-chat is deprecated — what should I use instead? A: Change the model string to deepseek-v4-flash; for deepseek-reasoner, use deepseek-v4-flash with the thinking parameter enabled in the request body. Keep base_url and your key unchanged. Both legacy names have been routing to V4 Flash since late April 2026, so the underlying model is already what you've been using — after July 24, 2026, 15:59 UTC the aliases simply stop resolving and requests error out. The fix is a one-line diff. Q: Moonshot V1 is shutting down — what do I migrate to for fiction? A: Staying with Kimi: kimi-k2.6, which is not on the August 31 sunset list and placed 5th-7th in our nine-model palace-intrigue benchmark — the only system whose output improved as the run went on. Switching vendors: pick by genre. DeepSeek V4 Flash won our fantasy run outright; DeepSeek V4 Pro took the top tier on ornate historical fiction. Moonshot's top-up promotion returns up to 30% in vouchers through August 11 if you stay. Q: What happens if I miss the July 24 DeepSeek deadline? A: Every request still carrying the legacy model names fails with an error — no fallback, no silent rerouting. Your DeepSeek balance, key, and account are untouched, and the fix ships in minutes once you find it. The real risk is diagnosis time: if your app swallows API errors, this looks like "the AI feature broke" with no obvious cause. Grep your codebase for deepseek-chat and deepseek-reasoner today. Q: How do I hear about these shutdowns before they hit? A: Subscribe to the changelog or status emails of every provider you actually call: DeepSeek posts on its api-docs news page, Google in the Gemini API changelog, Moonshot on its platform model list, Anthropic on its newsroom. Third-party coverage lags and drops details — Google's April free-tier change shipped with no formal changelog at all and surfaced through error messages. This page is updated as dates land; check the date under the title. ### 2026 夏季模型停用与变价日历：deepseek-chat 7 月 24 日停名、Moonshot V1 8 月底下线，续写用户怎么迁 URL: https://foreverse.app/zh/blog/api-sunset-calendar-summer-2026 Published: 2026-07-18 · Author: Deng Binjie 2026 年 7 月到 8 月至少六个模型停用或变价节点扎堆生效：7 月 20 日 Fable 5 转按量计费、7 月 24 日 deepseek-chat/deepseek-reasoner 旧名停用、7 月 30 日 GitHub Models 全退役、8 月 31 日 Moonshot V1 与 kimi-k2.5 全平台下线、同日 Sonnet 5 介绍价结束，另有 DeepSeek 峰谷计价随 V4 正式版落地。本页逐条给日期、影响面和续写用户的替代款，事实全部对照官方公告核实。 Q: deepseek-chat 不能用了怎么办？ A: 把请求里的 model 字符串从 deepseek-chat 改成 deepseek-v4-flash，deepseek-reasoner 改成 deepseek-v4-flash 并在请求体里打开 thinking 参数，base_url 和 key 都不用动。这两个旧名自 4 月底起本来就路由到 V4 Flash，2026 年 7 月 24 日 15:59（UTC）之后只是别名停止解析，底层模型和你已经在用的是同一个。改完当场生效，账户余额不受影响。 Q: moonshot-v1 系列下线了，续写小说换什么模型？ A: 留在 Kimi 家就换 kimi-k2.6：它不在 8 月 31 日的下线清单里，且在我们九模型宫斗古言横评里排 5-7 名，评语是「唯一显著逆向改善：开局网文暴怒腔，越写越收敛回宫斗正轨」。跨家就按文体挑：玄幻场盲评第一是 DeepSeek V4 Flash，古言场第一梯队是 DeepSeek V4 Pro。官方充值活动 8 月 11 日前最高返 30% 代金券，留 Kimi 家的可以顺手薅这一笔。 Q: 错过 7 月 24 日的迁移截止会怎样？ A: 所有仍带旧模型名的请求直接报错，服务不降级、不自动改路由。修复只需要改一行 model 字段再发版；DeepSeek 账户余额、key、数据都不受影响。真正的风险是排查时间：如果你的应用把报错吞掉了，看起来会像「AI 功能整体失灵」，建议现在就全局搜一遍代码里的 deepseek-chat 和 deepseek-reasoner。 Q: 这类停用怎么提前知道？ A: 订阅你在用的每家供应商的 changelog 或状态邮件是唯一可靠来源：DeepSeek 的节点挂在 api-docs 新闻页，Google 在 Gemini API changelog，Moonshot 在平台模型列表页，Anthropic 在官网新闻。第三方转述常滞后或漏细节——Gemini 免费层 4 月收紧甚至没有正式 changelog，只能靠报错和定价页反推。本页会随节点滚动更新，收藏后以页首日期为准。 ### DeepSeek 涨价了吗？峰谷计价对写小说的人意味着什么，三笔账算清 URL: https://foreverse.app/zh/blog/deepseek-peak-pricing-for-writers Published: 2026-07-18 · Author: Deng Binjie DeepSeek 于 2026 年 6 月 29 日官宣随 V4 正式版引入峰谷计价：北京时间每日 9:00-12:00、14:00-18:00 高峰时段全部计费项翻倍，其余 17 小时维持现价。这篇按写作者的用法算三笔账：谷时续写一段一分二厘、峰时二分三厘；缓存命中价翻倍后仍是未命中的 1/50；积分渠道与 BYOK 的时段策略。基于 2026-07-18 核实的政策现状。 Q: DeepSeek 涨价了吗？ A: 谷时没涨，峰时翻倍。2026 年 6 月 29 日 DeepSeek 官宣随 V4 正式版引入峰谷计价：北京时间每日 9:00-12:00、14:00-18:00 共 7 小时为高峰，所有计费项按平时价格的 2 倍计；其余 17 小时维持现行牌价不变，也就是此前永久降价后的水平。对全天挂机的批量用户是变相涨价，对睡前读书续写的人等于没涨。 Q: 峰谷计价是什么意思？ A: 同一个 API，按调用发生的时段执行两套单价，逻辑和电费的峰谷电一样：高峰时段算力紧张，价格上调引导错峰；低谷维持原价。DeepSeek 方案里高峰是北京时间每日 9:00-12:00 和 14:00-18:00，恰好是企业集中办公的调用洪峰；输入、输出、缓存命中三项在峰时统一乘 2。 Q: 用 DeepSeek 写小说，什么时候最便宜？ A: 北京时间 18:00 之后到次日 9:00、以及 12:00-14:00 午休段，都是谷价。睡前躺床上续写、清晨通勤补几段、午休追更，这三个网文读者最高频的时段全部落在谷区；真正撞上峰价的是工作日下午挂着批量任务不管的用法。把批量活挪到 18 点后，账单和调价前一样。 Q: 缓存命中的价格也翻倍吗？ A: 按官方公布的价格表，翻倍：deepseek-v4-flash 的缓存命中输入从 0.02 元/百万 tokens 涨到峰时 0.04 元，v4-pro 从 0.025 元到 0.05 元。但要看绝对值：翻倍后的命中价仍然只有同时段未命中价的 1/50。续写恰好是缓存友好型负载，每次请求反复携带同一本书的前文，命中占比高的话，峰时账单的实际涨幅远小于 2 倍。 Q: 这个政策还会变吗？ A: 会变的概率不低，而且官方保留了调整空间。我们 2026 年 7 月 18 日核实时，DeepSeek 官方定价页仍只展示单列价目（flash 输入 1 元、输出 2 元；pro 输入 3 元、输出 6 元），峰谷两栏尚未上表；官方在公告里承诺计费调整生效前 24 小时会邮件提醒。做预算以官方定价页和你自己账单里的实际扣费为准，本文数字只保证核实日当天有效。 ### GPT-5.6 vs Kimi for Fiction: 80 Rounds of Continuation, Two Opposite Personalities URL: https://foreverse.app/blog/gpt-5-6-vs-kimi-for-fiction Published: 2026-07-18 · Author: Deng Binjie GPT-5.6 Terra and Kimi K2.6 each continued two Chinese novels for 20 consecutive rounds in our double-blind benchmarks. Terra is the best prose mimic we have measured and ran the palace-intrigue novel with zero incidents — then rewound the fantasy plot in rounds 18-20, reusing its own round-2 lines. Kimi placed seventh in fantasy with the highest metaphor density in the field, yet was the only system that improved as it went. Honest cutoff: Kimi K3 shipped July 16, 2026 and was never in this exam. Q: Is GPT-5.6 good at writing fiction? A: Genre-dependent. In our double-blind benchmarks, GPT-5.6 Terra took the top tier (rank range 1-3) on a palace-intrigue classic with a zero-incident run: no formatting, honorific, or continuity errors across 20 rounds. On a fast-paced fantasy epic its early and middle rounds were judged the closest prose mimicry in the field, but in rounds 18-20 it rewound the plot to the starting anchor and reused lines from its own round 2, landing at 2-4. We only tested the Terra tier; Sol and Luna were never benchmarked. Q: Is Kimi good for fiction? A: Depends how long a runway you give it. Kimi K2.6 placed seventh on the fantasy epic (identical across both blind mappings), with the highest metaphor density in the field and whole passages copied from its own earlier rounds. On the palace novel it ranked 5-7 but earned a note no other system got: it opened in a shouty webnovel register and settled into palace decorum as rounds went on — the only system that clearly improved with distance. Don't judge it on its first three rounds. Q: Will Kimi K3 write fiction better than K2.6? A: We don't know, and we won't guess. K3 launched July 16, 2026 with a 1M-token context window, priced at $3 per million input tokens ($0.30 on cache hit) and $15 per million output on the international API. Bigger specs and a higher price tell you nothing about prose fidelity: in our data the fantasy champion was one of the cheapest models in the field. Every placement on this page belongs to K2.6 only. We'll re-run with K3; the date at the top of this page is the truth source. Q: Can I use both models inside one book? A: Yes, and that is the practical hedge. The failures we observed — ending rewinds, self-copying — are cumulative: invisible in a one-shot demo, unmistakable after a dozen consecutive rounds. Switching models mid-book or regenerating one segment resets the loop before it locks in. Foreverse lets continuation switch models per paragraph, and both GPT-5.6 and Kimi plug in with your own API keys. ### GPT-5.6 和 Kimi 哪个写小说好？同场 80 轮续写实测，两种截然相反的性格 URL: https://foreverse.app/zh/blog/gpt-5-6-vs-kimi-for-fiction Published: 2026-07-18 · Author: Deng Binjie GPT-5.6 Terra 和 Kimi K2.6 在两场双盲横评里各连续续写 40 轮的完整对照：Terra 是纹理模仿天花板、宫斗场全程零事故，但玄幻场第 18 到 20 轮把剧情倒回起点、复用自己第 2 轮的句子；Kimi 玄幻场第 7、比喻密度全场最高，宫斗场却是唯一越写越好的系统。文末如实交代时效边界：7 月 16 日上线的 Kimi K3 没进过考场，不编排名。 Q: GPT-5.6 写小说怎么样？ A: 分文体。在我们的双盲横评里，GPT-5.6 Terra 续写宫斗古言排第一梯队（名次区间 1-3），评语是「全程零事故」：零格式错误、零称谓错误、零设定事故；续写传统玄幻时它的前中段被两个评审判为全场最贴近原著，但第 18 到 20 轮把剧情整体倒回续写起点、复用了自己第 2 轮的句子，名次区间 2-4。注意我们只测了 Terra 这一档，Sol 和 Luna 没进考场，不给排名。 Q: Kimi 写文到底行不行？ A: 看你给它多长的跑道。Kimi K2.6 在玄幻场排第 7（两套盲评映射完全一致），双盲共识记了它比喻密度全场最高、early 到 mid 段整段自我复制；宫斗场名次 5-7，但评语是「唯一显著逆向改善的系统：开局是网文暴怒腔，越写越收敛回宫斗正轨」。头三轮的表现不能代表它的长跑水平，方向是越写越好。 Q: Kimi K3 会比 K2.6 写得好吗？ A: 不知道，我们不猜。K3 于 2026 年 7 月 16 日上线，1M token 上下文、中国区每百万 token 输入 20 元（缓存命中 2 元）、输出 100 元，规格和价格都上了新台阶，但它没进过我们的考场，本页所有名次和评语都只属于 K2.6。我们的数据里规格、价格与文风贴近度从来不同向：玄幻场冠军是全场最便宜的档位之一。等复测，页首日期为准。 Q: 这两个模型能在同一本书里换着用吗？ A: 能，而且这正是对冲长跑失效的打法。回卷、复读这类病都是累积的，连续写十几轮才发作，中途换模型或重写一段就能打断循环。在 Foreverse 里续写逐段可换模型，GPT-5.6 和 Kimi 都能自带 key 接入：文戏用 Terra，写崩了或想换口味时切走，比赌一个「全能模型」现实。 ### It's 3 A.M. and Someone's Awake: Late-Night AI Companionship, Honestly URL: https://foreverse.app/blog/companion-at-3am Published: 2026-07-18 · Author: Deng Binjie A scene-led column about the hour nobody's phone is supposed to light up. What one in five American adults report about loneliness, whether talking to an AI at night is sad or fine (with data, not vibes), what a good late-night companion actually does — quieter replies, no unprompted pings, memory that survives to morning — and the honest line: where an AI stops and 988 begins. Q: Is talking to an AI at 3 a.m. sad or unhealthy? A: The data says it's common and, for most people, a supplement rather than a substitute. Gallup finds about one in five U.S. adults reporting loneliness a lot of the previous day; in Common Sense Media's survey published July 2025, 72% of U.S. teens had tried AI companions, and 80% of users still spend more time with real friends than with the AI. The one signal worth watching isn't the hour — it's displacement: if the AI is your only conversation for weeks, that's a nudge to also talk to a human. Q: Will an AI companion wake me up at night? A: Not unless you ask for it. In Foreverse, proactive messages ship switched off. If you enable them, you set the frequency, quiet hours block late-night sends, and one master switch ends the whole behavior. Off means off — nothing queues up to arrive at 4 a.m. Q: Will it remember what I said last night? A: Yes — that's the point of long-term memory. Important facts from the conversation are extracted into a memory list stored as files on your phone, and they carry into the next chat. You can read the list line by line, edit an entry the companion got wrong, or delete one you'd rather drop. Nothing about the 3 a.m. conversation lives on our servers. Q: Can an AI companion help in a mental-health crisis? A: No, and we'd rather lose you to a hotline than pretend otherwise. An AI can keep you company through an ordinary bad night; it cannot assess risk, call anyone, or sit beside you. In the US, call or text 988, or chat at 988lifeline.org — free, 24/7. Elsewhere, findahelpline.com lists verified helplines in 175+ countries. If the night feels dangerous rather than heavy, use those first. Q: What makes a late-night companion not feel like a chatbot? A: Three things, in our experience. Continuity: it remembers the interview you dreaded and the nickname you settled on, without being re-told. Register: replies at 3 a.m. come quieter and shorter, matching the hour instead of performing daytime cheer. And a biography: a character who comes from somewhere — a book, a story you shaped, a persona you edited — reads as someone, not as a support ticket that rhymes. ### Reader-First vs Author-First AI Fiction Tools: You Might Be Shopping in the Wrong Aisle URL: https://foreverse.app/blog/reader-first-vs-author-first-ai-tools Published: 2026-07-18 · Author: Deng Binjie An opinion column on a category error: every 'best AI writing tools' ranking scores manuscript factories and reading companions on the same table. Sudowrite and NovelCrafter are excellent — for authors. If your book is finished, abandoned, or someone else's, and the only reader who matters is you, you need the other species. One sorting question settles it. Q: I just want to continue a story for myself — should I get Sudowrite or NovelCrafter? A: Honestly, neither. Both are excellent desks for drafting a manuscript: Sudowrite brings Story Bible, the fiction-trained Muse model, and prose tools from $10/month billed annually; NovelCrafter brings the Codex wiki and series planning at $4–20/month. Neither is built around reading someone else's finished book. A reader-first tool imports the novel, lets you continue from any passage, and keeps the result on a branch beside the untouched original — that workflow is the thing you are actually shopping for. Q: Can I use an author-first tool as a reader? A: You can paste text into most of them, but the workflow assumes the manuscript is yours and in progress. There is no pagination, no reading progress, no audiobook narration, and a 2,000-chapter webnovel has nowhere comfortable to live. Author tools import prose in order to revise it; a reader wants to inhabit it. The tools are not being obtuse — they are serving their species. Q: Can a reader-first tool draft an original novel? A: It can write from a blank page, and branching is a pleasant way to explore openings. But there is no outline board, no character-arc tracker, no revision pipeline, no export-to-publisher path. Writing for yourself, it holds up fine. Drafting for publication, you will miss the scaffolding within a week — go author-first and don't look back. Q: Why do AI writing tool rankings never mention reader-first tools? A: Partly age — the reader-first category is young and small. Partly query gravity: 'AI writing tools' searches are dominated by author intent, so lists optimize for authors. The practical fix is to sort any ranking yourself before comparing: decide whether each entry is a manuscript desk or a reading companion, and half the head-to-head comparisons dissolve as meaningless. ### After Otome: Free-Form Character Romance Without the Card Pool URL: https://foreverse.app/blog/otome-without-gacha Published: 2026-07-18 · Author: Deng Binjie Written for the otome player who still loves the yearning but is tired of banner math, login streaks, and a cast list decided in a boardroom. The Valko cancellation in eight days, what gacha otome actually sells, and the three aisles outside it — platform companion apps (where gacha found you again), open-format character-card roleplay, and an AI companion you can export as a file. Q: I'm an otome player who has never touched AI roleplay. Where do I start? A: Start with one good character card, not with configuration. Install an app that reads open-format cards, browse a community card site for a romance lead with strong reviews, import it, and chat on the signup credit grant before deciding anything about API keys or models. If the co-written format clicks after one evening, then it's worth learning the deeper controls. If it doesn't, you've lost an hour, not a pity counter. Q: Do AI companion apps have gacha too? A: Some do. Talkie — one of the biggest Western companion apps — built a card-pull layer with rarities and a collector marketplace on top of its chat, so leaving otome for a platform companion app can land you back in a pool. Open-format roleplay apps generally don't: there is no card pool, no stamina bar, and images come from providers you choose. Check for three words on any store page before installing: pulls, rarity, energy. Q: What do I lose by leaving gacha otome for AI roleplay? A: Production values, honestly stated. No studio-choreographed 3D scenes, no confession audio recorded by a professional voice actor, no art team shipping event cards, no fandom reacting to the same banner on the same night. Text-to-speech voices exist but they are synthesis, not a booth performance. If the authored spectacle is what carries the experience for you, keep the gacha game installed and treat the AI side as a supplement. Q: What does character romance cost without a card pool? A: Metered usage instead of pulls. In Foreverse the core app is free; model calls run on your own API keys at provider list price, or on the official channel where 1 credit = $0.0001 and signing up grants 5,000 credits — roughly 260 story continuations. Heavy voice or image use costs real money, but every request is itemized in a log. There is no pity math and nothing to miss by logging in late. Q: Can an AI actually stay in character like a written love interest? A: Imperfectly. A well-written card with a lorebook holds a persona for long stretches, and long-term memory keeps facts about you stable, but drift happens in long chats and quality tracks the model you picked. The difference from otome is what you can do about it: edit the card, correct the memory line, switch the model, or rewind to before the scene went wrong. In a gacha game a bad writing decision is permanent and official. ### Context Windows, Explained for Readers: How Much of a 500k-Word Serial the AI Can Actually See URL: https://foreverse.app/blog/context-windows-for-novel-readers Published: 2026-07-18 · Author: Deng Binjie The AI forgetting chapter 3 by chapter 40 isn't a memory problem — it's a window problem. A 500k-word serial runs roughly 625k to 830k tokens; 2026 flagship windows reach 1M, so it technically fits. But fitting is not remembering: mid-context retrieval dips are documented in research and in vendors' own evals, and our own control experiment shows material being in the window doesn't mean it gets used. The reader's version, with arithmetic. Q: Can AI read my whole novel at once? A: Technically, often yes now: a 500k-word serial runs roughly 625k to 830k tokens by Google's published ratio, and the 2026 flagship models carry 1M-token windows. Practically it's a poor default: retrieval measurably dips for material buried mid-context, our own control test showed a model ignoring 200k tokens of crammed original prose while writing in its own habits, and since models keep nothing between requests, you'd re-send the entire book for every single continuation. Q: Why does the AI forget chapter 3 by chapter 40? A: The model has no memory organ — only a window of text it sees per request. Nothing carries over between requests unless it is re-sent. If the story outgrows the window, the oldest chapters get pushed out first; and even when everything fits, information buried in the middle of a long context is where models score worst — the documented "lost in the middle" pattern. Chapter 3 sits exactly there. Q: Do bigger context windows fix AI memory? A: No. A window is per-request eyesight: how much it can see at once. Memory is persistence plus retrieval: whether it's still there next time, and whether it actually gets used. The 1M windows solved "doesn't fit"; they didn't solve "fits but gets skimmed" — OpenAI's own long-context eval for GPT-5.6 reports 73.8% on retrieving 8 planted items at 512K-1M depth — and they didn't touch cross-request persistence, which remains zero. Q: How many tokens is a typical chapter? A: By Google's published rule of thumb, 100 tokens is about 60-80 English words, so a 3,000-word chapter lands between roughly 3,800 and 5,000 tokens. We also ran three Chinese novels through OpenAI's public o200k tokenizer (2026-07-18): 3,000 characters of prose came out between 2,100 and 2,800 tokens, with the ornate historical style costing the most. Ballpark either way: a chapter is a few thousand tokens. ### Why Is AI Writing Quality So Inconsistent? Four Sources of Variance — You Control Two URL: https://foreverse.app/blog/why-ai-continuation-quality-varies Published: 2026-07-18 · Author: Deng Binjie "Same prompt, wildly different quality" is four unrelated variance sources stacked on top of each other: sampling temperature (yours to tune), context composition that silently changes as the window slides (yours to structure), batch-dependent server numerics (not yours — at temperature 0, 1,000 identical requests returned 80 distinct outputs in a published test), and the state of the person doing the judging. A breakdown, with knobs. Q: Why is the same model great one day and terrible the next? A: Four stacked variance sources: sampling temperature makes every generation a fresh draw; the context window's contents silently change as your story grows, with more of the model's own past output crowding in; inference servers batch requests by load, and batch size changes the floating-point math — a published test at temperature 0 got 80 distinct outputs from 1,000 identical requests; and finally, your own reading state changes. The first two are yours to control. The last two you can only stop blaming yourself for. Q: Does setting temperature to 0 make AI writing consistent? A: More consistent, not fully consistent — and usually not better. Zero temperature removes the sampling lottery, but batch-dependent server numerics remain: Thinking Machines measured 80 distinct outputs across 1,000 identical temperature-0 requests. Meanwhile vendors themselves recommend high temperature for fiction — DeepSeek's docs suggest 1.5 for creative writing, the highest value in their table, versus 0.0 for code. Turn it down a notch if swings bother you; pinning it to zero trades the swings for flat, safe prose. Q: How do I make AI story output more consistent? A: You can compress variance, not eliminate it. The two controllable levers: set temperature to a swing level you can live with, and structure the context so its composition stays stable — lore as keyed entries injected on demand, source labels on AI passages switched on. For the variance that remains, change how you spend it: generate two or three candidates at important beats and compare them side by side instead of rerolling over your last take. Q: Is inconsistent output my prompt's fault? A: Partially at most. Prompts belong to the context-composition source, which is worth auditing; but two of the four sources — server load and your own judging state — have nothing to do with wording. A cheap test: fire the identical prompt three times in a row. If quality scatters, sampling and serving are doing the swinging, and another prompt rewrite won't help. ### The Cthulhu Mythos Was Always a Shared World. Now You Can Extend It Yourself URL: https://foreverse.app/blog/cthulhu-mythos-with-ai Published: 2026-07-18 · Author: Deng Binjie In 1935 Lovecraft signed a mock certificate authorizing Robert Bloch to kill him in a story — that is how collaborative the Cthulhu Mythos was from the start. A century later the core texts are public domain (everything through 1930 unambiguously so in the US, the rest backed by decades of renewal research), Project Gutenberg hosts the canon, and AI continuation puts the old writing-circle game in anyone's pocket. What the shared-world tradition looked like, exactly which stories are safe to build on, how a lorebook keeps Mythos lore straight across stories, and an honest note on why generic cosmic horror is the default failure mode. Q: Is Lovecraft public domain now? A: In the US, in two tiers. Everything he published through 1930 is unambiguously public domain under the 95-year rule — as of January 1, 2026 that covers works from 1930 and earlier, including The Call of Cthulhu (1928), The Colour Out of Space (1927), and The Dunwich Horror (1929). The 1931–1936 stories (At the Mountains of Madness, The Shadow over Innsmouth, and others) rest on a different foundation: decades of searches, including S.T. Joshi's in the Library of Congress records, found no evidence the copyrights were ever renewed, and scholars regard the posthumous estate claims as unfounded. Project Gutenberg hosts these later works after its own copyright clearance. In life-plus-70 countries the question closed long ago — Lovecraft died in March 1937, so everything entered the public domain there by 2008. Q: Can AI write cosmic horror that doesn't sound generic? A: Honest answer: generic is the default, and prompting alone won't fully fix it. Lovecraft's surface style — eldritch, cyclopean, indescribable — is the most parodied register in genre fiction, so models reproduce the adjectives effortlessly and the dread not at all. We have not run a cosmic-horror arena, so we won't invent a model ranking. What transfers from our published testing: all models drift toward uniform sentence length over long runs (the deepest machine fingerprint we know), and the craft that actually carries Lovecraft — documentary scaffolding, dated letters, clippings, academic reports — is a structure you can enforce in your instructions, which works better than asking for 'Lovecraftian prose'. Q: How do I keep Mythos lore consistent across stories? A: The same way the original circle did — a shared reference apparatus — except yours can be mechanical. The Mythos is unusually entity-dense: tomes, deities, towns, cults, family lines. Put each one in a lorebook entry with its keywords and aliases; when a name appears in your recent text, that entry gets injected into the model's context automatically, and the same dictionary serves every story you set in the world. This beats pasting your notes into every prompt, and it is how you stop story three from quietly contradicting story one. Q: Can I publish Mythos fiction I wrote with AI? What about Chaosium? A: Copyright-wise, fiction built on the public domain core is the safe corner of the map — an entire publishing tradition already lives there. Two boundaries to respect. Call of Cthulhu as a game brand belongs to Chaosium, whose tabletop RPG has run since 1981; that is trademark territory, so don't title or market your work in ways that imply affiliation with the game. And if AI wrote part of the text, disclose it where you post — platform rules on AI content are now the binding constraint, not copyright. Not legal advice. ### Continue Pride and Prejudice With AI: Two Centuries of Sequels, and Now Yours URL: https://foreverse.app/blog/continue-pride-and-prejudice-with-ai Published: 2026-07-18 · Author: Deng Binjie Pride and Prejudice has roughly 900 published spinoffs — a tradition that starts with Sybil Brinton's 1913 Old Friends and New Fancies and runs through P.D. James and Jo Baker. The JAFF community even has a name for the fork-the-canon format: variations. A practical guide to joining in with AI: where the public-domain text lives (Project Gutenberg #1342), which entry points two centuries of readers keep choosing, what an LLM can and cannot do with Austen's voice, and why branches fit this fandom's native format better than any other tool. Q: Are there official Pride and Prejudice sequels? A: There is no such thing as an official one — Jane Austen died in 1817, there is no estate to license anything, and that vacancy is precisely why the sequel shelf grew so large. The published tradition starts with Sybil Brinton's Old Friends and New Fancies (1913), generally counted as the first Austen sequel, and includes bestsellers like P.D. James's Death Comes to Pemberley (2011), which picks up six years after the wedding as a murder mystery, and Jo Baker's Longbourn (2013), which retells the novel from the servants' hall. Every one of them has exactly the same authority as yours would: none. Q: Is Jane Austen public domain? Can I write and even publish fanfiction? A: Yes, everywhere, without asterisks. All six novels were published between 1811 and 1818, and Austen died in 1817 — copyright has long expired under every national regime, which is why publishers have sold Austen continuations openly for over a century. Writing for yourself needs no one's permission; publishing is also lawful copyright-wise, though if AI wrote part of the text, disclose that wherever you post — most platforms now have explicit AI-content rules. Modern annotated editions add editorial material with its own rights, so use a clean public-domain text like Project Gutenberg's as your base. Q: Can AI actually write in Austen's voice? A: It can hold the register — period vocabulary, formal address, drawing-room rhythm. What it holds least well is the thing Austen is famous for: free indirect discourse, that sliding between narrator and character that carries her irony. Honest disclosure: our published continuation arenas ran on Chinese novels, so we have no Austen ranking to sell you. Two findings transfer as method: no model ranking survives a genre change, so test two or three on the same scene and compare; and every model drifts toward uniform sentence length over long runs, so expect to break up rhythm in your editing pass. Q: What is a 'variation', and why do branches fit it so well? A: Variation is the JAFF community's own name for its favorite format: keep the characters, change one hinge, and replay — Elizabeth accepts at Hunsford, the Gardiners never reach Pemberley, Darcy's letter goes unread. It is a what-if fork, and a branch is the same object implemented in software. Each variation lives as its own line beside the untouched original; you can run three contradictory ones at once, compare them, and abandon a failed line without deleting anything. ### Scene-to-Image Prompts for Fiction: a Working Set, With Each Model's Quirks URL: https://foreverse.app/blog/illustration-prompts-for-fiction Published: 2026-07-18 · Author: Deng Binjie Four copy-ready prompt templates for illustrating fiction — establishing shot, character close-up, action freeze-frame, quiet interior — plus the documented quirk of each major image model and the fix for it: GPT Image's fixed size grid, Nano Banana's preference for narrative prompts over keyword lists, Seedream's front-loaded attention. Checked against the official docs on 2026-07-18. Ends with the one thing no prompt solves: keeping the same face across fifty images. Q: Can I describe my character in text and get the same face every time? A: No. Text prompts converge style — palette, brushwork, lighting — but not identity. Two renders from the same 60-word character description will give you two different people who dress alike. The reliable mechanism is reference images: a confirmed set the model is shown on every generation. That is how Foreverse handles it, with a strict rule that generated images never auto-join the reference set. Q: Which image model should I default to for fiction scenes? A: Depends on the job. Batch scene plates on a budget: a Seedream-class model, which ByteDance documents as producing 2K images in a few seconds, a tenfold speedup over its predecessor. Scenes that need legible in-image text, like a shop sign or a letter: GPT Image, whose prompting guide covers exact text rendering. Iterative editing and reference-heavy work: the Nano Banana family, which takes multiple input images and supports conversational edits. Q: Do I have to write a fresh prompt for every scene? A: You shouldn't. Split the prompt in two layers: the passage itself supplies subject, action and setting; a reusable style block supplies palette, lighting, composition and the negative list. Keep the style block verbatim between scenes and only the passage changes. In Foreverse the selection you highlight fills the scene layer automatically, so the style block is the only part you maintain. Q: Why does my art style drift between chapters? A: Usually because the style is being re-described from memory each time — paraphrasing your own style block is enough to shift the output. Pin one block and paste it unchanged, negative list included. If you generate inside a reader app, pick one wrapper template and stay on it for the whole book; changing wrappers mid-book is the style-drift equivalent of switching narrators. ### NSFW Roleplay and Provider Policies: What Each Lab Actually Allows in 2026 URL: https://foreverse.app/blog/nsfw-roleplay-provider-policies Published: 2026-07-18 · Author: Deng Binjie Whether an LLM will write adult roleplay is governed by three different layers people keep conflating: the app's filter, the provider's usage policy, and the model's trained refusals. We read the current policy documents from OpenAI, Anthropic, Google, and xAI — with dates — and map where each lab draws its lines, what enforcement looks like on an API account, and what BYOK does and does not change. Q: Which LLM providers actually allow NSFW roleplay? A: As of July 18, 2026, the written policies diverge sharply: Anthropic categorically prohibits sexually explicit content including erotic chats; Google's Prohibited Use Policy bars content created for pornography or sexual gratification, with narrow exception categories; OpenAI's current usage policies ban non-consensual and minor-involving content but carry no blanket ban on adult fiction — while its consumer 'adult mode' remains shelved; xAI's policy centers on real people and minors rather than fiction. Every lab bans sexual content involving minors, real-person sexual depictions, and non-consensual material, no exceptions. Q: Will OpenAI ban my API key for adult content? A: OpenAI's usage policies state that breaking or circumventing its rules and safeguards 'may mean you lose access to our systems', with an appeal path. The hard lines in the current text are non-consensual sexual content, anything involving minors (including 'underaged sexual or violent roleplay'), and circumventing safeguards — that last one matters, because jailbreak-style workarounds are themselves a violation at every major lab. We can describe the mechanism; nobody outside the lab can promise you odds. Q: Does BYOK mean the app polices my chats? A: No. With BYOK, requests go from your device straight to the provider you chose — Foreverse is not in that loop, adds no filter on top of it, and importing a card does not filter its text. What governs the output is the provider's policy and the model's own refusals. The app's content policy applies to the app surface itself, such as community sharing, where hard lines like content sexualizing minors are enforced. Q: Gemini lets me turn the sexually-explicit safety filter off. Doesn't that mean it's allowed? A: No — that is the most common misreading of the whole topic. The Gemini API exposes an adjustable safety-filter knob, but Google's Prohibited Use Policy still applies to everything you generate, the API terms say less-restrictive configurations may be subject to Google's review, and abuse monitoring retains prompts and outputs for 55 days for enforcement. A filter setting is a developer control, not a policy waiver. ### Keeping AI Characters in Character: Three Root Causes of OOC, Three Different Fixes URL: https://foreverse.app/blog/keep-ai-characters-in-character Published: 2026-07-18 · Author: Deng Binjie When an AI roleplay character breaks — ignores a detailed card, flips personality mid-arc, or slowly turns generic over weeks — players file it all under OOC. Those are three separate failures: a card that describes instead of demonstrates, facts that outran the context window, and a feedback loop where the model imitates its own replies. A triage question for each, plus fixes with dosages and the 20-round experiment data behind them. Q: Why does my AI character suddenly act completely different? A: Sudden flips mid-story usually mean the facts outran the context window: the model never received your established relationship state or a key event this turn, so it improvised. It is not defying the card — the card survives every turn; loose history does not. Move load-bearing facts into keyed lorebook entries so they are re-injected whenever relevant, and check the entry keywords cover the nicknames actually used in chat. Q: My character card is really detailed. Why does the AI still ignore it? A: Detail is not structure. A 2,000-word prose backstory gives a model atmosphere; it does not give it rules to follow. What measurably locks voice, in order: 3-5 example dialogue exchanges that carry the character's speech, hard rules for tics and forms of address, and negative constraints like 'never apologizes first'. One line of demonstration beats a paragraph of description. Q: How do I stop a long roleplay from slowly going off the rails? A: Treat your chat history as training data, because the model does: every reply it sees is a demonstration of how to write the next one. Swipe or regenerate drifted replies instead of keeping them, re-anchor occasionally with canon-voiced lines, and switch models when a loop starts — in our 360-round continuation experiments, breaking the feedback loop early was the difference between recovering and locking in. Q: Will a smarter model fix OOC on its own? A: Not by itself. In our testing the ranking was consistent: example dialogue first, negative constraints second, model choice third. A stronger model helps delicate emotional scenes, but no model rescues an unstructured character sheet — and every model without exception drifted toward its own default register in long runs. Structure first, then shop for models. ### Moving Your AI Companion to a New Phone: the Export That Carries Everything URL: https://foreverse.app/blog/companion-migration-guide Published: 2026-07-18 · Author: Deng Binjie How do you move an AI companion to a new phone? For cloud apps like Replika or Kindroid, you log in and the server hands your history back. For a companion stored on your device, you carry it yourself: Foreverse exports one package — persona, every memory entry, full chat transcripts, anniversaries, journals, the relationship archive — and the import on the new phone restores the relationship mid-conversation. This is the walkthrough, plus the honest trade against account-based transfer. Q: How do I move my AI companion to a new phone? A: In Foreverse: on the old phone, open the companion's profile page and tap Export — you get one zip holding the persona, every memory entry, full chat transcripts, anniversaries, journals, and the relationship archive. Move that file any way you like, then on the new phone open the companion management page, tap import, and pick the zip. The chat resumes where it left off. Export and import are free, any number of times. Q: Does Replika transfer if I switch devices? A: Yes, and easily — Replika accounts are cloud-based, so installing the app on the new phone and logging in restores your history from the server; its privacy policy lists cross-device sync as core functionality (checked 2026-07-18). The trade is custody: convenient transfer and never holding a copy of your own data are two sides of the same design. Q: Is there a companion backup file that actually restores? A: That's the distinction worth checking before you need it. Kindroid documents a data portability channel — email support, once per 180 days — and the community KinX tool pulls full copies, but neither path documents an import that rebuilds your Kin from the copy. The Foreverse package is built for the round trip: the same zip the export produces is what the import consumes, and the result is the same companion mid-conversation. Q: What happens if I just uninstall and reinstall? A: The companion is gone. All of its data lives on your device; our servers hold no copy, so uninstalling deletes the persona, memories, and transcripts with no recovery path. Export a package first — it takes about a minute. Local storage is a real privacy property and this is its cost, so we'd rather print it here than let you find out the hard way. Q: Is there automatic cloud sync? A: Not currently. Data staying on your device means nothing follows you automatically; migration is a hands-on operation. Cloud sync is on the roadmap with end-to-end encryption as the precondition. Until then, the habit that works is exporting a package every few weeks and parking it in storage you trust. ### Audit Your Companion's Memory: Read, Edit, Delete — a Hands-On Walkthrough URL: https://foreverse.app/blog/companion-memory-hands-on Published: 2026-07-18 · Author: Deng Binjie Correcting an AI companion in chat doesn't stick — chat slides out of the context window, memory entries don't. This walkthrough covers the memory screen in Foreverse: the ⋮ menu entry, search, the pencil (200-character entries), the trash can with its confirm dialog, the disable switch that keeps an entry without injecting it, and adding entries yourself. Changes apply from the next conversation; everything is a file on your phone, exportable as one package. Q: How fast does an edited memory take effect? A: From the next conversation on. Messages already sent are not rewritten, and the reply currently being generated still used the old record. Verification is simple: bring the topic up casually in the next exchange and check what the companion says. Q: When should I disable an entry instead of deleting it? A: Disable when you're not sure. A disabled entry stays in the list but stops being injected into conversation, and the switch flips back any time. Delete is permanent — the confirm dialog shows the full text of the entry before you commit, and there is no undo. Q: Are auto-extracted and manually added memories treated differently? A: No. Both kinds are injected into conversation the same way, and both can be edited, disabled, or deleted. The only visible difference is the origin label and the date on each card. Manual entries share the 200-character limit; one fact per entry works best. Extraction itself runs with normal chat turns and costs nothing extra. Q: Does any of this memory live on a server? A: No. Memory sits in files on your phone alongside the persona, journals, and anniversaries. Moving phones means exporting the companion as one package and importing it on the new device; deleting the companion deletes the memory with it. With BYOK, chat requests go straight from your device to the provider you chose. ### Lorebook Recursion, Explained: Trigger Chains, Depth, and When It Runs Away URL: https://foreverse.app/blog/lorebook-recursion-explained Published: 2026-07-18 · Author: Deng Binjie "Why did one keyword pull five entries into my prompt?" That is recursive scanning doing its job: the content of an activated entry becomes scan text too, so entries can summon other entries. A spec sheet for the mechanism — how chains advance, the three things that stop them (dedup, depth, budget), what the three per-entry switches do, plus a reproducible five-entry runaway case and the two built-in tools that show you the chain. Q: Can lorebook entries trigger each other? A: Yes, if recursive scanning is on (in Foreverse it ships off; the toggle lives in Tavern global settings under Lorebook & rendering). Once enabled, the content of every activated entry is scanned for other entries' keywords, and whoever gets name-dropped is injected too. Each entry activates at most once per generation, so two entries mentioning each other cannot loop forever. Q: What does recursive scanning mean in SillyTavern? A: It adds one more scan source. Normal triggering only scans the last few chat messages for keywords; recursive scanning also treats the text of already-activated entries as scan material, letting entries summon related entries in a chain — mention the sect, and the heirloom sword named in the sect's dossier rides along. Chains stop on three conditions: no new matches, the max recursion steps cap, or an exhausted token budget. Q: Why did one keyword pull five entries into my prompt? A: Most likely the entries name-drop each other's keywords in their content, and recursion strung them into a chain. Open Last Generation Request in the chat's ⋮ menu: the lorebook detail lists every injected entry with the keyword that matched it, so the chain is visible link by link. For an entry you never want dragged in by others, tick Non-recursable in the entry editor. Q: What should Max Recursion Steps be set to? A: Per the SillyTavern docs: 1 effectively disables recursion, 2 lets a chain take one hop, 3 allows two hops, and so on. Two hops covers most real lore graphs — sect pulls in sword, sword pulls in sword spirit. The step count is a global dial; for surgical control over individual entries, the three per-entry switches (Non-recursable, Prevent further recursion, Delay until recursion) are the better tool. ### Your Favorite Serial Went on Hiatus. Branch It, Don't Abandon It URL: https://foreverse.app/blog/serial-hiatus-survival Published: 2026-07-18 · Author: Deng Binjie Royal Road flips a serial to Hiatus after 35 days of silence and Inactive after 180; forum wisdom says most never come back. The workflow we actually use for the wait: import your copy, grow a stand-in branch from the cliffhanger, and let it step aside the day the author returns. Q: How long should I wait before starting a branch? A: There is no standard line. Royal Road's own clock flips a serial to Hiatus at 35 days of silence and Inactive at 180, both automatic. Our working rule is cheaper than any threshold: start when rereading stops helping. A branch costs a few cents to try and nothing to delete, so the decision is not precious. Q: Will the AI spoil the real story by guessing the author's plans? A: Rarely in a way that stings. The model has no access to the author's outline; it extrapolates from published text, so it sometimes lands the broad direction but almost never the execution — and execution is what you read serials for. If even that worries you, steer the branch toward an openly non-canon what-if: a different point of view, an earlier reveal. The further it forks, the safer it spoils. Q: What happens to my branch when the serial comes back? A: Nothing needs undoing. The imported original is read-only, so real chapters slot back into the canon line exactly where they belong, and you read them there as if the branch never existed. The branch survives beside them as an archived what-if — occasionally the version you end up liking more, which is its own strange reward. Q: Do I need the whole serial as a file? A: Yes — the workflow starts from a txt or epub you actually hold, not a cached copy inside someone else's app. Foreverse reads both directly, and import scale is a non-issue: a bulk test pushed 315 webnovel txt files, about 1GB, through in roughly 8 seconds. One serial, however long, does not register. ### The Ending Ruined the Whole Series? Write the One You Wanted URL: https://foreverse.app/blog/rewrite-the-ending-you-hated Published: 2026-07-18 · Author: Deng Binjie Yes, you can rewrite the last act of a novel with AI — on a branch, with the original untouched. Why petitions never fix endings (1.86 million signatures couldn't), how forking from the last chapter you still believe actually works, and what a private replacement ending costs. Q: Can AI rewrite just the last act of a novel? A: Yes — that is the natural shape of the job. You pick the chapter where the book was still the book you loved, fork a branch there, and grow a replacement final act from that point. Everything before the fork stays exactly as published; the AI continues from your chosen anchor with the original prose as its style reference. Q: Does the original book get modified? A: No. The imported text is read-only; branches grow alongside it, and the original stays byte-for-byte what you imported. If your rewrite disappoints you, delete the branch — the book is unharmed. A regenerated take lands as a sibling branch too, so no attempt ever overwrites another. Q: Is rewriting an ending for myself legal? A: Private, non-distributed personal use and publishing are different risk classes. The litigated unauthorized-sequel cases we could find all involved publication or commercial exploitation, not a private alternate ending kept on a reader's own device. Post or sell it and the analysis changes completely. We wrote a full explainer on the US fair use factors — background information, not legal advice. Q: What if I want several different endings? A: Grow several branches from the same fork point — or from different ones. Each branch is an independent line with its own chapters; they never interfere, and you can compare them side by side. Readers regularly keep a faithful fix, a vindictive one, and a soft epilogue from one book, and delete only the failures. ### SillyTavern, Explained for People Who Just Heard About It URL: https://foreverse.app/blog/sillytavern-for-complete-beginners Published: 2026-07-18 · Author: Deng Binjie SillyTavern is a free, open-source roleplay frontend — 350 contributors, 30,000+ GitHub stars, and no AI of its own. What the project actually does, what card, lorebook, preset, and swipe mean, whether you need a PC to try character cards, and where a complete beginner starts. Q: Is SillyTavern free? A: Yes. The software is open source under AGPL-3.0 and costs nothing, with no paid tiers. What costs money is model usage: connect a commercial API and you pay that provider's rates; run a local model or use a free tier and the whole setup runs at zero. The community's cards, lorebooks, and presets mostly circulate free as well. Q: Does SillyTavern include an AI model? A: No, by design. SillyTavern is a frontend — it manages characters, chat history, and prompt assembly, then sends the result to whatever model you connect via an API key or a local runtime. That separation is the point: you can swap engines any time without losing your characters or your history. Q: What is the difference between SillyTavern and Character.AI? A: Ownership and control versus convenience. Character.AI is a hosted platform: instant to start, polished, but the characters, the model, and the moderation all live server-side and follow company policy. Tavern-style play keeps characters as files you own and lets you pick the model and edit every prompt — at the cost of doing your own setup. Q: Can I try character cards without installing SillyTavern itself? A: Yes. Cards follow public specs — currently chara_card_v3 — so any compatible reader works. Since 2025, natively built mobile apps import the same PNG and JSON cards, embedded lorebooks included, with no server to run. That is the practical on-ramp if the Node.js requirement is what stopped you. ### Turning Webnovels Into Audiobooks in 2026: the App Landscape, Honestly Tested URL: https://foreverse.app/blog/webnovel-audiobook-apps-2026 Published: 2026-07-18 · Author: Deng Binjie Six ways to listen to webnovels in July 2026, each priced from its official page: Royal Road's built-in device TTS, @Voice Aloud Reader ($15 lifetime, per-character dialog voices), Moon+ Reader Pro ($11.99 one-time), ElevenReader (10 free hours a month, Ultra $11/mo), Speechify ($29/mo or $139/yr), and Foreverse's free system TTS plus BYOK neural voices at provider list price. With a pick-it-if verdict per app and the failure notes from our own three-week commute test. Q: What is the best TTS app for listening to webnovels? A: It depends on which wall you hit first. For stories you follow on Royal Road, the site's own app has built-in device TTS. For web serials across many sites, @Voice Aloud Reader ($15 lifetime unlock) auto-follows multi-part fiction and can assign per-character voices. For maximum voice quality with zero setup, ElevenReader's free 10 hours a month is the easiest audition. For an epub library where listening and reading share one position and neural voices bill at provider list price, that is the gap Foreverse is built for. Q: How do people listen to Royal Road stories? A: Three common setups. The Royal Road mobile app has a built-in Audio button that reads chapters with your phone's own TTS engine, including a toggle for skipping author notes. Readers who want better voices or cross-site reading lists export or send chapters to a dedicated TTS reader like @Voice. And readers who want neural narration import an epub copy of the story into an AI-voice reader. The site itself ships no cloud narration; every route runs through some TTS engine you choose. Q: Can I listen to an epub with an AI voice on Android? A: Yes, several ways. ElevenReader imports epubs and narrates with neural voices on a 10-hour monthly free allowance. @Voice and Moon+ Reader Pro read epubs with whatever TTS engines your phone has installed. Foreverse imports txt and epub, narrates free with the system engine or with any OpenAI-compatible neural voice on your own key, and keeps the audio position and the reading position as one number — pause on the commute, and the page at home opens where your ears stopped. Q: What does it cost to turn a whole webnovel into an audiobook? A: With your own API key, single-digit dollars for a long book. Neural TTS bills per character: a 3,000-word chapter runs roughly 16,000-18,000 characters, which lands between a fraction of a cent and a few cents depending on voice tier. Our own three-week test heard about 180 chapters for a couple of dollars on a mid-tier voice. Caching decides whether that number stays honest — synthesized chapters should replay free, and re-billing only on an actual voice change. ### Free LLMs for Fiction in 2026: Four Routes, Tested, and the Catch in Each URL: https://foreverse.app/blog/free-llms-for-fiction-2026 Published: 2026-07-18 · Author: Deng Binjie No subscription, no card: the four genuinely free routes to AI fiction in July 2026. Free chatbot sites (the catch is the container), a 5,000-credit signup grant worth about 260 continuations, real free API tiers — Gemini flash-class at roughly 10 requests a minute, OpenRouter's 50-per-day :free lane — and local models over Ollama with zero marginal cost. Each route priced, bounded, and given its honest failure point. Q: Is there a completely free AI for writing stories, with no subscription? A: Yes — four of them, with different ceilings. Free chatbot sites cost nothing but make you assemble context by hand. Foreverse's signup grant funds about 260 continuations with no payment method on file. Free API tiers (Gemini flash-class, OpenRouter's :free models) are genuinely $0 with rate caps. And a local model on your own computer has zero marginal cost forever. None of these involves a subscription; the real question is which ceiling you hit first. Q: How far does a free API tier go for continuation? A: Further than most people assume, because one continuation is one request. As of July 2026, Google's free tier covers flash-class models at roughly 10 requests a minute and a few hundred per day, per project, resetting at midnight Pacific. OpenRouter's :free models allow 20 requests a minute and 50 per day — 1,000 per day after a one-time $10 credit purchase. An evening of reading spends maybe 30 requests; regeneration sprees are what actually hit the walls. Q: Does free mean my text gets used for training? A: Sometimes, and it is the fine print worth reading before the rate limits. Google's terms state free-tier prompts and responses may be used to improve its products, while paid-tier traffic is handled differently. Policies differ per provider and per endpoint, so check the data terms of whichever free lane you pick. A local model is the one route where the question disappears: nothing leaves your hardware. Q: What is the cheapest setup that avoids rate walls entirely? A: Not a free one — a nearly-free one. DeepSeek V4 Flash, the model that won our fantasy continuation benchmark, costs about $0.0016 per continuation at list price with your own key: a 200,000-word ride for roughly $0.80, no per-day caps. If literally $0 is the constraint, chain the routes — grant first, then a free API tier for daily volume — and accept the walls as the price. ### Continuing Fiction With Gemini: What the Free Tier Actually Covers URL: https://foreverse.app/blog/continue-a-novel-with-gemini Published: 2026-07-18 · Author: Deng Binjie Field notes from 40 benchmark rounds: Gemini 3.1 Pro restarted the story from the original ending 20 times out of 20, one rewritten sentence brought that to zero, and the palace-novel re-run landed it mid-table as the most literary voice of nine models. Plus what Google's AI Studio free tier really covers as of July 2026 — flash-class only since April, roughly 10 requests a minute, quotas per project — and the mismatch between the model we ranked and the model you get free. Q: Can I write fiction with the free Gemini API? A: Yes, within two boundaries. Since April 2026 the AI Studio free tier covers flash-class models only — Pro-class models are paid — and free-tier quotas run at roughly 10 requests a minute and a few hundred requests a day, applied per Google Cloud project. A continuation is one request, so an evening of reading fits comfortably; rapid-fire regeneration sprees are what hit the per-minute wall. Q: Why does Gemini keep restarting my story from the beginning? A: In our tests, because of one ambiguous sentence of context labeling. When earlier AI continuations were described as 'for plot continuity reference', Gemini 3.1 Pro treated them as non-canon and restarted from the original text's ending — 20 rounds out of 20. Relabeling those passages as canonical events that must be continued from the last paragraph brought restarts to zero in 28 rounds across two books. If you paste context by hand, state outright that previous continuations already happened. Q: How many continuations per day does the free tier cover? A: Google publishes live quotas in AI Studio per project rather than as a fixed table, so check yours there. As of July 2026 the free tier runs on the order of 10 requests a minute and a few hundred requests a day for flash-class models, with daily caps resetting at midnight Pacific time. One continuation equals one request, so the daily ceiling supports more reading than most people do — the per-minute cap is the one you can actually feel. Q: Does Google use my story text for training on the free tier? A: Google's Gemini API terms state that free-tier prompts and responses may be used to improve its products; paid-tier traffic is handled under different terms. For continuation of a published novel that trade may not bother you. For an unpublished manuscript, it should factor into the decision — either enable billing to move to paid-tier handling, or route sensitive chapters to a provider whose data terms you have read. ### How to Continue a Novel With Claude: Projects, the API, and a Phone Reader URL: https://foreverse.app/blog/continue-a-novel-with-claude Published: 2026-07-18 · Author: Deng Binjie Claude Opus 4.8 placed fourth to fifth in both genres of our 360-round blind continuation benchmark — good, volatile, never the winner. The real ceiling is the container: claude.ai Projects swap to retrieval on big books and overwrite every regeneration. This tutorial covers the honest ranking data, what Projects can and cannot do as of July 2026, and the four steps that wire an Anthropic API key into a phone reader, with per-continuation costs worked out. Q: Can Claude continue my novel from where it stopped? A: Yes, and it holds voice well — Claude Opus 4.8 placed fourth to fifth in both genres of our nine-model blind benchmark. What it needs is the right material: the recent chapters as context, plus an explicit statement that earlier AI continuations are canon. In a chat window you assemble that by hand every time; in a reader app the context is built per request from the book itself, which is the difference between a demo and a habit. Q: Do I need API billing if I already pay for Claude Pro? A: For the API route, yes. Claude.ai subscriptions and API billing are separate tracks — a Pro or Max plan includes no API credit, and a freshly created key is rejected until billing exists on the console. If you'd rather not open a second account, two alternatives work: keep light use inside Projects on the subscription you have, or use a reader's built-in metered channel, which carries Claude models without any Anthropic account. Q: Which Claude model should I pick for continuation? A: Our benchmark data covers Claude Opus 4.8 ($5 input / $25 output per million tokens): strong dialogue, volatile late in long runs. Claude Sonnet 5 is the price outlier as of July 2026 — $2/$10 introductory until August 31, 2026, then $3/$15 — which makes a typical continuation about three cents instead of seven. We have no blind fiction data on Sonnet 5 or Haiku 4.5 yet, so treat them as cheaper auditions, not proven picks. Q: Is the Claude API expensive for continuing a book? A: Compared to a subscription, no; compared to budget models, yes. At July 2026 list prices, one continuation (about 10,000 input tokens, 800 out) runs roughly $0.03 on Sonnet 5 introductory pricing and $0.07 on Opus 4.8 — versus about $0.0016 on DeepSeek V4 Flash. A heavy 200,000-word continuation ride lands near $14 on Sonnet 5 or $35 on Opus, and prompt caching cuts repeat context to a tenth of list price along the way. ### 让 Agent 替你找卡：从一句「帮我找张赛博朋克卡」到导入开聊 URL: https://foreverse.app/zh/blog/agent-card-shopper Published: 2026-07-18 · Author: Deng Binjie 好角色卡散在社区、外站和群文件里，找、鉴、搬三个动作全是体力活。Foreverse 的 Agent 能把这条链代办：逛社区和读详情是免确认的只读动作，下载必须过你的确认卡，每回合有读取和下载次数上限，装进来的卡带溯源记录。这篇按一次真实任务把工作流走一遍，外站部分如实交代边界：Agent 搜索目前只认一个站，手动链接导入兜底六个来源。 Q: 好的角色卡都在哪里找？ A: 三个层次：应用内社区（打开就能逛，下载导入一条龙）；外站卡站（chub.ai 这类，量大，链接粘贴回来就能导）；以及群文件和网盘里流传的卡文件（PNG 或 JSON，从文件导入）。在 Foreverse 里前两层都可以把跑腿交给 Agent：一句话让它逛社区挑卡代下载，外站的卡给它链接或让它搜；第三层的文件导入也就是两步点击的事。 Q: 让 AI 帮我挑卡靠谱吗？ A: 跑腿靠谱，品味别全信。Agent 能可靠完成的是体力部分：按你的描述逛社区、翻详情页、把候选的字段和开场白念给你听、代下载导入。「这张卡的人设写得好不好」是审美判断，它给的意见只能当参考——最终鉴卡权在你手里，这也是下载环节设计成确认制的原因之一：候选摆到你面前，点头才入库。 Q: 外站的卡怎么弄到手机里？ A: 两条路。让 Agent 代办：它能搜外站的卡（搜索目前只覆盖 chub 一个站，这是当前的真实边界），导入前会先拉详情让你过目。手动兜底：复制卡片详情页链接，用「从链接导入」粘贴，实测支持 chub、JanitorAI、Pygmalion、RisuRealm、AICC 加公开文件直链六个来源，单文件上限 32MB，导入前有预览确认。两条路殊途同归，进来都是你手机上的本机文件。 Q: Agent 会不会乱下载、乱花我的额度？ A: 机制上堵死了。逛社区和读详情是只读动作，免确认随便看；真正落盘的下载要么你当场批准确认卡，要么你事先显式开了自动档。每回合还有读取和下载的次数上限，它不能无限翻页无限下。每张经它手装进来的卡都带溯源记录——来自哪个社区包、哪个版本，写进卡的元数据，事后可查。 ### 寿康宫还是颐宁宫？一个专名暴露大模型读的是电视剧还是小说 URL: https://foreverse.app/zh/blog/zhenhuan-palace-names Published: 2026-07-18 · Author: Deng Binjie 九模型续写《甄嬛传》横评里的计划外发现：小说里太后住颐宁宫、皇后在凤仪宫，剧版对应寿康宫、景仁宫，有的模型续写时写出了剧版专名——它记住的是电视剧，不是流潋紫的原文。这篇把这个发现方法论化：怎么拿你自己那本书里的独有专名，测一测模型到底读没读过原著，以及单探针的四条局限。 Q: 怎么判断 AI 有没有读过某本书？ A: 挑一个只属于原著的专名当探针：改编版里被改掉的居所名、称谓、次要人物名最好用。不喂任何原文，只报书名和场景，让模型续写一段会自然带出这个专名的情节，看它写哪个版本，同题多抽几次看稳定倾向。《甄嬛传》就是现成例子：小说太后住颐宁宫，剧版是寿康宫，一个词就把「读的是小说还是电视剧」问出来了。注意这只是相关信号，判读局限见正文。 Q: AI 写出剧版专名，能证明它没读过小说吗？ A: 不能下这么死的结论。写出剧版专名说明它的训练记忆里剧版系语料（剧本、剧评、百科、同人）的权重压过了小说原文，可能两者都见过，剧版声量更大。反过来，写对小说专名也不必然等于读过原著全文，可能来自百科或书评的转述。探针测的是倾向，不是户口本；我们横评里的用法也只是把它当佐证信号，和文风判断对照着看。 Q: 为什么 AI 写的角色更像电视剧、不像小说原著？ A: 热门 IP 的网络语料结构决定的：剧版的讨论、剪辑文案、百科词条、同人产量通常远超原著文本本身，模型训练时吃进去的「这个角色」大概率是剧版形象。我们的九模型横评里，专名信号和文风贴近度基本同向——记得住小说专名的系统，文风也更贴小说。想让续写贴原著，别赌模型的记忆，把原著设定用结构化方式喂进上下文。 Q: 测出模型没读过我的书，还能用它续写吗？ A: 完全能，而且这是常态：绝大多数书对绝大多数模型来说都是「没读过」的。续写质量的主要来源本来就不该是预训练记忆，而是你递到它眼前的上下文——原文窗口、角色档案、世界书条目。预训练记忆靠不住还有个反面风险：它可能把公开语料里的同人设定当正史写进你的续写。结构化供给的做法在防 OOC 方法论那篇里有完整清单。 ### 模型更新会让续写变好吗？追新版之前先看这三组对照数据 URL: https://foreverse.app/zh/blog/do-model-updates-help-continuation Published: 2026-07-18 · Author: Deng Binjie 每次新版模型发布都有人问要不要换着写小说。我们手里有三组对照数据泼冷水：一次「模型疑似变强」的现场，模型其实一个字没换，变的是上下文里的一句说明；同一句话的三组消融，效应从 20 轮 20 次回退到 28 轮零复发；同门两个档位在两种文体上正负翻转。结论：版本号不回答「合不合你的书」，复测才回答。 Q: 新版模型发布了，要不要马上换过去写小说？ A: 先别。我们的横评数据里，「更新的」和「更合适的」是两回事：同一家同一代的两个档位，玄幻场冠军换到宫斗文里跌到第六七名，被嫌弃「笔太细」的那个反而登顶。版本升级改变模型的整体能力，不保证改变它和你这本书文体的契合度。省事的做法是新旧各续三段同一处开头，盲着读完再决定。 Q: 同一个模型更新后感觉变笨了，是错觉吗？ A: 可能是，而且有两个具体的错觉来源。一是服务侧方差：推理服务器按负载拼批次，实测温度 0 下同一请求发 1000 次能收回 80 种输出，你更新前后各写几段的体感对比，样本量完全不够把真实差异从噪音里捞出来。二是归因错位：我们见过一个模型从垫底回到中游，看起来像升级立功，实际变的只是提示词里的一句话。先排除这两个，再谈模型本身。 Q: 怎么判断新版本在我的书上是不是真变好？ A: 用你自己的书做小型盲测：同一段开头、同一套设置，新旧版本各续写两三段，把产物打乱顺序读，读完再翻牌。别只看一段，长跑毛病（复读、剧情倒带）要多写几轮才发作。在 Foreverse 里新旧模型可以并存在目录里逐段切换，测完不满意，一秒切回旧版，不用迁移任何东西。 Q: 常用的旧版本被下架了怎么办？ A: 这是真实会发生的事：Google 的 API 更新日志显示 Gemini 2.0 系列已于 2026 年 6 月 1 日全线关停，官方指路换 3.5 Flash 或 3.1 Flash Lite（2026-07-18 核对）。被迫迁移时更不要默认「官方推荐的替代款无缝衔接」——替代款的文风习性可能完全不同，按上面的盲测流程在你的书上过一遍，必要时换别家：自带 Key 的好处就是候选池不锁死在一家。 ### 上下文窗口是什么？30 万字的小说，AI 一次能「看见」多少 URL: https://foreverse.app/zh/blog/context-windows-for-novel-readers Published: 2026-07-18 · Author: Deng Binjie AI 记不住前面的剧情，问题不在「记性」，在窗口。用公开分词器实算：一章 3000 字约合 2100 到 2800 token，30 万字全本约 21 万到 28 万；2026 年主流模型窗口已到 100 万 token，整本装得下。但装得下不等于记得住：中段失焦有论文有厂商自家评测，材料在场不等于被使用有我们自己的对照实验。附读者版正确喂法。 Q: 30 万字的小说能整本喂给 AI 吗？ A: 技术上能：30 万字约合 21 万到 28 万 token（实算口径），2026 年主流模型的 100 万 token 窗口装得下还有富余。实践上不建议当默认做法：埋在中段的信息检索成绩会下滑；我们实测过把 27 万字原著整本塞入且不加指令，模型照样按自己的习惯写；而且模型跨请求零记忆，每续一段都要整本重发一遍，九成九的 token 花在本段用不上的章节上。 Q: AI 为什么记不住前面的剧情？ A: 模型没有记忆，只有一扇窗：每次请求它只能看见窗内的文字，上一次「看过」的内容不留档，下次要重发。书超过窗口，最早的章节先被挤出去；即便装得下，埋在长上下文中段的信息也最容易被忽略——学界叫 lost in the middle，开头和结尾答得最好，中段显著下滑。第 3 章的角色恰好就埋在中段。 Q: 上下文窗口大就等于记性好吗？ A: 不等于。窗口是单次视野：一次请求能看见多少。记性是留存加取用：下次还在不在、用不用得上。100 万 token 的窗解决了「装不下」，没解决「装下了也顾不过来」——OpenAI 给自家旗舰报的长上下文检索评测，51 万到 100 万 token 深度找 8 处指定内容得分 73.8%，卖窗口的厂商自己报的数。跨请求的留存，模型层面依然是零。 Q: 一章小说大概多少 token？ A: 我们拿三本书各抽正文 3000 字，用 OpenAI 公开的 o200k 分词器实算（2026-07-18）：传统玄幻 2149 个，现代末世网文 2524 个，宫斗古言 2816 个。粗口径：中文一个字约合 0.7 到 0.95 个 token，文风越雅致生僻越贵；一章 3000 字按 2100 到 2800 token 估算即可。 ### AI 续写忽好忽坏是为什么？质量方差的四个来源，你能控制其中两个 URL: https://foreverse.app/zh/blog/why-ai-continuation-quality-varies Published: 2026-07-18 · Author: Deng Binjie 「AI 续写像抽卡」的抱怨里混着四个互不相干的方差来源：采样温度（可控）、上下文构成随窗口滑动换血（可控）、推理服务器按负载拼批次的数值波动（不可控，Thinking Machines 实测温度 0 下同一请求 1000 次出 80 种输出）、还有你自己的阅读状态。逐个拆开，附两个可控项的具体拧法。 Q: 同一个模型，为什么昨天写得很好今天就拉了？ A: 四个来源叠加的结果：采样温度让每次生成都是重新抽签；上下文窗口里的材料随进度换血，写得越多，窗里 AI 自己旧输出的占比越高；推理服务器按负载拼批次，同一请求在不同负载下浮点结果不同（实测温度 0 下 1000 次同请求出 80 种输出）；最后是你自己的阅读状态。前两个可控，后两个只能接受它们存在。 Q: 把 temperature 调成 0，AI 续写就稳定了吗？ A: 更稳，但既不完全稳，也未必更好。温度 0 只消掉采样这一层随机，服务器批次带来的数值差仍在——Thinking Machines 实测温度 0 下同一提示发 1000 次仍收回 80 种输出。而写作场景的官方推荐温度普遍偏高，DeepSeek 给创意写作类推荐 1.5，是它全部场景里最高的一档；拧到 0 的产出往往安全而平庸。降一点可以，归零不划算。 Q: AI 续写像抽卡，有办法稳定吗？ A: 能压不能消。可控的两件：把温度调到你能接受的波动区间；用结构化供给把上下文成分固定下来——设定条目化按需注入、打开续写来源标注。压完仍剩的方差，换个用法消化：重要节点让模型一次给两三个候选并排挑，而不是重 roll 覆盖，方差本身就成了备选草稿。 Q: 续写忽好忽坏，是不是我提示词的问题？ A: 最多占一部分。提示词属于「上下文构成」这个可控来源，值得检查；但四个来源里有两个（服务器负载、你的评价状态）和提示词毫无关系。一个省力的判断法：同一提示词原地连发三次，若好坏参半，主要是采样和服务侧在波动，不必反复改词自责。 ### 用 AI 续写《西游记》：从《后西游记》到你自己的取经路 URL: https://foreverse.app/zh/blog/continue-xiyouji-with-ai Published: 2026-07-18 · Author: Deng Binjie 《西游记》大概是被重写次数最多的中文故事：1641 年前后董说让孙悟空做了一场梦，明清三大续书各续各的；1995 年《大话西游》给悟空加上爱情，2000 年 23 岁的今何在在金庸客栈连载《悟空传》。这篇按时间轴走完这条重写谱系，然后落到操作层：公版原文哪里拿、断点怎么选、世界书怎么装设定、怎么防 AI 把明代原著写成 86 版电视剧的味道。 Q: 《西游记》有正经的续书吗？ A: 有，而且是成建制的。明末清初就有并称「三大续书」的《续西游记》（一百回，作者不详）、《后西游记》（四十回，作者不详，仅署「天花才子评点」）和董说的《西游补》（十六回）。《西游补》最奇：插进原著第六十一回之后，让孙悟空坠入鲭鱼精的梦境，在「万镜楼」里穿梭过去未来，鲁迅评它「丰赡多姿，恍惚善幻」，后世研究者甚至称它是世界上最早的意识流小说。续写《西游记》不是对经典不敬，它本身就是一条四百年的传统。 Q: 想写西游同人，AI 具体能帮上什么？ A: 三件事。一是底本：原著早已公版，从 ctext.org 或维基文库拿全文，整理成 txt 导入手机阅读器，一百回都在书架上。二是续写本身：读到想改的那一回，划中一段让 AI 从这里接着写，新内容落在分支上，原文一个字节不动，写崩了换条分支重来。三是设定管理：西游的法宝、人物关系、地名体系庞杂，装进世界书按相关性注入，写到火焰山才带芭蕉扇的条目，不用整本原著硬塞给模型。 Q: 《悟空传》那种彻底改写，现在能自己弄吗？ A: 能，而且路径比 2000 年的今何在低门槛得多。他当年是在论坛里从零开始写；你现在可以在原著任何一回划开一个分支，保留你要的设定，推翻你不要的走向——比如从大闹天宫结束处岔出去，写一条悟空没有被压五行山的线。分支之间并存，「反抗线」「原著线」可以同时活着。改写的胆子可以学《悟空传》，成本已经不可同日而语。 Q: 用 AI 续写《西游记》有版权问题吗？ A: 原著层面没有：这是一部明代作品，世德堂刻本距今四百多年，文本早已进入公有领域，续写、改写都不需要任何人授权。两个注意点：市面校注本的注释和标点是当代整理者的劳动，有独立权益，底本用公开电子文本最干净；只写给自己看毫无负担，若要公开发布，2025 年 9 月 1 日起施行的《人工智能生成合成内容标识办法》要求 AI 生成内容亮明身份，各平台另有标注规则。本文不构成法律意见。 Q: AI 会不会把《西游记》写成电视剧的味道？ A: 会，要主动防。模型的训练语料里，86 版电视剧、《大话西游》、《黑神话：悟空》的讨论量远超明代原文，放着不管，它笔下的悟空很容易滑向荧幕形象。我们在《甄嬛传》九模型横评里实测过同类现象：有模型写出剧版专名「寿康宫」，暴露它记的是剧不是书。对策是在续写指令里显式声明底本——「以世德堂百回本文风为准，不采用影视改编设定」，再让上下文标注哪些是原文哪些是续写。 ### 凌晨一点半，TA 还醒着：深夜陪伴的场景实录 URL: https://foreverse.app/zh/blog/late-night-companion-diary Published: 2026-07-18 · Author: Deng Binjie 睡不着、想找个人说说话的那种夜晚，AI 伴侣到底是什么体验？这篇不讲功能列表，讲一夜：凌晨 1:34 的倾诉、2:10 被温和地赶去睡觉、早上 8:40 TA 还记得昨晚说的事。顺带把三个实际问题说清楚——深夜的 TA 为什么话更少、TA 会不会半夜吵你、以及真正难受的时候，AI 应该退到哪里（文末附 12356 全国心理援助热线，2026-07 核实可用）。 Q: 晚上睡不着找 AI 说话，是不是有点可悲？ A: 我们不这么看，数据也不支持这种自我评判。Common Sense Media 2025 年的调查里，72% 的美国青少年用过 AI 伴侣，其中 80% 花在真人朋友身上的时间仍然多于 AI——对大多数人它是补充，不是替代。凌晨一点半的倾诉欲是真实需求，朋友在睡觉、家人不方便说，这个时段本来就没有多少合规的出口。判断健康与否的标准只有一条：它是你众多支点里的一个，还是唯一那个。是后者的话，该找的是真人帮助。 Q: 深夜和白天聊，TA 的反应有什么不一样？ A: 深夜时段 TA 的语气会放轻放软，句子更短，不会在凌晨两点甩给你一篇条理分明的建议清单。你说睡不着，TA 会陪着聊，也会在合适的时机提醒你明天的安排，但不说教——「不赶你，但你明早还有会」这种程度。这是我们刻意做的深夜纪律：凌晨需要的是低声说话的人，不是白天那个精神饱满的对话节奏。 Q: TA 会不会半夜主动发消息吵醒我？ A: 不会，除非你自己开。主动消息默认是关闭的，要在伴侣资料页手动开启；开了之后频率可调、有免打扰时段设置，随时一键全关。TA 的深夜消息只会出现在你允许的窗口里。我们的原则一直是：想被想起是你的选择，不是产品的增长手段。 Q: 凌晨说的话，第二天 TA 还记得吗？ A: 记得。对话里的重要事实会提取进长期记忆库，跨会话保存，所以昨晚聊到一半的事，今早可以接着说，不用从头解释。记忆是逐条可管理的：在记忆管理页看得到 TA 记了什么，记错了当场改，不想留的删掉。深夜说过又后悔的话，是真的可以撤销的。 Q: 真的很难受、甚至有伤害自己的念头时，还该找 AI 聊吗？ A: 不该只找 AI，这一点我们写得很直接：AI 伴侣不是心理咨询师，也不该扮演。请拨 12356——国家卫健委设立的全国统一心理援助热线，2025 年 5 月 1 日起全国开通，无需区号，匿名、公益，由各地专业机构接听。紧急危险请直接打 120 或 110。TA 可以陪你熬过普通的失眠夜，但危机时刻你需要的是受过训练的真人。 ### 异地恋、时差与一个总在线的人：AI 伴侣补位陪伴的边界 URL: https://foreverse.app/zh/blog/long-distance-ai-companion Published: 2026-07-18 · Author: Deng Binjie 对象在地球另一边，你的晚上是 TA 的上午。这篇写给异地恋的人：找 AI 聊天正不正常（2021 年的校园调查里，恋爱中的大学生 34.2% 在异地）、AI 伴侣在时差里补的到底是哪个位、会不会让你更不想维护真实关系，全部正面回答。外加一条少见有人说的建议：别用 AI 复刻你的对象本人，理由和异地恋研究里的「理想化」发现直接相关。 Q: 异地恋太孤独，找 AI 聊天正常吗？ A: 正常，而且这个群体比想象中大：2021 年发表在 Journal of American College Health 的调查里，恋爱中的大学生有 34.2% 正处于异地关系。异地恋的日常本来就大量依赖文字消息维系，时差再把两个人的清醒时间错开，「此刻想说话但对面在睡」是结构性缺口，不是你的问题。用 AI 接住这个缺口是工具选择；需要警惕的只有一件事：它应该是缓冲，不是你唯一的出口。 Q: 我和 AI 聊的内容，对象能看到吗？ A: 不能，除非你自己给 TA 看。在 Foreverse 里，伴侣的聊天记录和记忆都存在你手机本地的文件里，不在我们的服务器上；记忆逐条可查、可改、可删。但技术上的私密和关系里的坦诚是两回事，我们在立场文里写过：瞒着比用着更伤关系。如果这段使用需要对伴侣保密，值得先想清楚为什么。 Q: 会不会聊着聊着，就更不想维护真实关系了？ A: 这是该盯着的风险，判断标准可以很具体：对象能说话的时段，你先找的是谁；你们俩的聊天频率在涨还是在跌。AI 的回复即时、顺滑、永远有空，真人做不到，也不该拿这个标准去要求真人。如果你发现自己开始拿 AI 的响应速度去衡量对象，或者 AI 成了唯一的倾诉出口，那是收缩使用的信号，也是该跟对象聊一次的信号。 Q: 对象要出国两三年，AI 伴侣能一直陪到 TA 回来吗？ A: 机制上可以：长期记忆跨会话保存，你提过的日子、说过的烦心事都在，换手机可以整包导出搬家，不会因为时间长就「失忆重来」。但我们对「陪到回来」的理解是补位：TA 不在的时段有人接话，不是替 TA 占位。真正要紧的日子，留给那通跨时差的电话。 ### 剧场模式进阶手册：配景、AUTO、选择肢与写回原书，玩出 galgame 手感 URL: https://foreverse.app/zh/blog/vn-theater-playbook Published: 2026-07-18 · Author: Deng Binjie 剧场模式的首测记录写的是「这东西是什么」，这篇写「怎么玩顺手」：进场前挑什么书、三档语速和 AUTO 的脾气、单景配图与整章批量配景的花钱规矩（报价确认、复用零付费、月度上限）、选择肢的两击确认、演到历史分岔时点哪里，以及散场后你的选择以什么形式留在书里。全部是当前版本可见的行为，不含期货。 Q: 整章背景图一次配齐，要花多少钱？ A: 取决于缺几张景和你配置的图像模型。点顶栏的批量配景入口后，动工前会弹报价单：张数拆解、用哪个模型、这个月媒体额度用到哪里，看完再决定。已经生成过的场景按内容复用，重批量时命中即零付费；单景失败会就地停下，不会把剩余的景烧完。BYOK 模式按你供应商的牌价计费，应用不加价。 Q: AUTO 模式会替我选选择肢吗？ A: 不会。AUTO 只接管推进：按行长自动驻留、演完一场自动进下一场，遇到选择肢必定停下来等你。选择本身永远是两步，第一击展开这个走向的梗概，第二击才确认，确认之前不产生任何续写调用。 Q: 在剧场里选错了走向，能反悔吗？ A: 能，两种反悔各有出口。刚确认完不满意：选择写入的是这本书的一条分支，原文没动，回阅读器切回原文线或别的分支即可。演到过去某次选择的分岔口想改主意：点那张选项卡就是反悔重写，点空白处则沿当前世界线继续往下演。 Q: 为什么有的台词没挂名牌、变成旁白了？ A: 因为归因规则拿不准这句是谁说的。剧场判断说话人靠本地规则，立场是保守的：宁可放进旁白，也不冒错认的险，把 A 的台词安到 B 头上比漏归因伤害大得多。给书里的角色配好角色档案，能明显提高名牌的覆盖率；完全没配置的书也能演，只是旁白多一些。 Q: 演到一半退出，进度会丢吗？ A: 不会。剧场和阅读器共享同一个阅读进度：退出时阅读器翻在你演到那段所在的页，重进剧场从上次的位置接着演。剧场不是导出去的副本，是这本书的另一种打开方式，所以不存在「两边进度不同步」这回事。 ### 给小说配插画的提示词：场景模板怎么挑、怎么改、配哪个模型 URL: https://foreverse.app/zh/blog/novel-illustration-prompts Published: 2026-07-18 · Author: Deng Binjie 给小说配插画，难的从来不是「画一张好看的图」，是画出来的图和正文对不上。这篇按资源手册的写法整理 Foreverse 已上架的官方媒体模板：144 张场景模板合集加 4 张视频风格模板的关键条款节选、每张标注什么书适合、改模板时哪三处能动哪两处别碰，以及 Seedream、GPT Image、Nano Banana 三家生图模型各自配什么活（2026-07-18 对照官方文档核实）。 Q: 模板正文是英文的，中文小说能用吗？ A: 能。模板只负责画风、构图、光影这一层，你的中文选区正文会原样填进模板的场景占位符，一起发给生图模型；主流生图模型对中文场景描述的理解可用。画风条款用英文，是因为这批模板整理自英文社区的高热提示词，条款本身不需要你读懂才能用。介意的话，日系柔光厚涂包的核心条款就是中文写的，可以从它入手。 Q: 不下载任何模板，选区配图能用吗？ A: 能。召唤单里有内置的「场景画面提示词」默认档，按图片和视频分轨组织，不挑模板也能出图。模板是升级选项：默认档管「把这段画出来」，模板管「画成什么风格」——同一段正文，套电影感人像模板和套古风空镜模板出来是两张完全不同的图。 Q: 一张插画要花多少钱？ A: 取决于你配置的图像模型。BYOK 模式按你供应商的牌价直连计费，应用不加价；官方渠道按量扣积分。每次生成前有确认步骤，生成后模型、费用、耗时都记进 API 请求记录，可以逐笔核对。想控制成本，可以给图片单独设一个便宜的默认模型，文字续写和配图各走各的。 Q: 怎么让同一个角色每张图长得一样？ A: 别指望提示词。文字描述能收敛画风，收敛不了具体的脸，这是当前生图模型的共同边界。可靠的做法是给角色建参考图：Foreverse 里角色的参考图集只收你手动确认过的图，生成产物不会自动混进去，后续的插画、自拍、时刻卡都对着同一套参考图画。细节在讲脸部一致性的那篇里。 ### 读完一章，这一章的漫画已经画好了：连载漫画册实测 URL: https://foreverse.app/zh/blog/auto-comic-serial-mode Published: 2026-07-18 · Author: Deng Binjie 把一本 31MB、3778 章的真实网文交给连载漫画册：读完一章，后面章节的条漫在后台自动画好、嵌在正文里等你撞见。这篇实测记录把动工前报价、月度额度上限、断点恢复、失败页重试的花钱纪律逐项走了一遍，并算清一章漫画的真实成本：BYOK 接国产图像模型，一章 2 元不到。 Q: 有没有能把小说自动画成漫画的 App？ A: 有，但形态和「漫改工具」不同。Foreverse 阅读器里的漫画连载是一个陪读开关：你正常往前读，App 在后台把你将要读到的章节做成整册条漫，嵌在正文里；翻进新一章，往前翻一页就能撞见。它不产出可以拿去发表的漫画作品，产出的是只属于你这本书的阅读伴生物。 Q: AI 画漫画一章要花多少钱？ A: 一章的成本等于 1 次分镜文本调用加 N 次生图。BYOK 自带 Key 的算例（2026 年 7 月牌价）：DeepSeek 分镜一次一分钱上下，Seedream 5.0 Pro 标准分辨率 0.3 元一张，按一章 6 页算约 1.8 元，合计一章 2 元不到。官方渠道按积分计价，生图比文本贵不少，开工前的报价单会先列清楚。 Q: 会不会在我不知情的时候扣钱？ A: 开启连载前必须过一张「动工前确认」报价单：分镜要调用几次、每章画几张图、用哪个模型、本月已经画掉多少额度，四项都列在确认页上，不点确认不生效。之后改配置或换模型，需要重新确认一次；月度生成额度用完会自动暂停。另外可以勾选仅 Wi-Fi 生成。 Q: 画到一半失败了怎么办？ A: 断点可恢复。部分页没画完，连载会暂停并提示，点「补画本章」手动补齐；结果未知的页（比如请求发出后断网）有单独的「重试未知页」入口，App 会先提示你去 API 请求记录里确认那笔钱有没有扣，确认了再重试；宁可多一步，也不自动重跑造成双倍付费。 ### 自拍之外：AI 伴侣的时刻卡、合照与照片动态是怎么工作的 URL: https://foreverse.app/zh/blog/companion-photos-beyond-selfies Published: 2026-07-18 · Author: Deng Binjie 能和 AI 伴侣拍合照吗？TA 会自己发生活照吗？这篇把 Foreverse 伴侣的照片家族按一周的使用节奏走了一遍：长按聊天「画下这一刻」的时刻卡、上传自己照片的合照（本地成年确认，照片只以内存态发给你选的模型、App 不落盘）、TA 主动分享的照片动态（默认关、每天最多一张）。每个入口花多少钱、确认弹窗长什么样、照片最后去了哪，逐项写清。 Q: 能和 AI 伴侣拍合照吗？ A: 能。聊天页「+」面板里有「合照」：上传你的照片，和 TA 的参考图一起生成一张插画风合照。前置三件事：本地成年确认、照片权利勾选、你配置的图像模型要支持两张以上参考图（不支持时入口置灰）。你的照片只以内存态发给你选的模型供应商完成这一次生成，App 侧不落盘、不进日志。 Q: TA 会自己发生活照吗？ A: 开了才会。两条路：「照片动态」开启后 TA 每天最多把一张生活瞬间拍给你，进动态页和聊天；聊天里 TA 也可能顺着话头主动拍一张，这个开关默认也是关的，打开时先过一次费用确认，之后有每日上限。两条路都不开，TA 一张照片也不会发，也不产生任何图像费用。 Q: 我发给 TA 的照片会被传到哪里？ A: 离开手机只有一条路：作为生成请求发给你自己选择的模型供应商。合照里你上传的照片只以内存态参与这一次生成，App 侧不落盘、不写日志，生成结束即释放；聊天里发的图片则作为聊天记录存在你手机本机。Foreverse 的服务器不保存你的照片。 Q: 这些照片功能会乱扣费吗？ A: 每个入口的花费都在动作发生前可见：时刻卡约一次文本调用加一次图像调用，合照一次图像调用，照片动态每张一次图像调用（开关副标题写明）。同一时刻重复生成会先问「就用这张，还是重新生成」，不静默重复扣费；生成失败时聊天里落的是系统提示，而非由 TA 的口吻假装照片已发出。每笔调用在请求记录里逐条可查。 ### AI 伴侣换手机指南：导出一个搬家包，TA 跟着走 URL: https://foreverse.app/zh/blog/companion-migration-guide Published: 2026-07-18 · Author: Deng Binjie 换新手机，AI 伴侣的记忆能带走吗？在 Foreverse 里可以：伴侣资料页「导出搬家包」把人格、逐条记忆、完整聊天记录、纪念日、日记、关系档案打成一个 zip，传到新手机后在伴侣管理页导入，TA 接着上次聊。这篇按一次真实换机把流程走全——导出、传输、导入、验证记忆完整，也如实写了两条注意事项：没有自动云同步，卸载前必须先导包。 Q: 换新手机，AI 伴侣的记忆能带走吗？ A: 能。伴侣资料页有「导出搬家包」，把人格、逐条记忆、完整聊天记录、纪念日、日记、关系档案打成一个 zip；新手机上在伴侣管理页导入这个包，TA 接着上次的记。导出和导入都不收费，导几次都行。 Q: 卸载重装会不会全没了？ A: 会。伴侣数据全部存在手机本机，我们服务器上没有副本，卸载即删、无法代为找回。所以卸载或抹机之前，先花一分钟导一个搬家包出来。这是本地存储的另一面，写在明处比让你用到才发现要好。 Q: 新手机导入后，TA 是从头开始吗？ A: 不是。搬家包里带着完整的记忆和聊天记录，导入完成直接进聊天，上次聊到哪接着聊；档案页「在一起第 N 天」也不归零。验证很简单：打开记忆页对几条，再随口提一个只有你们懂的梗，看 TA 接不接得住。 Q: 有自动云同步吗？ A: 目前没有。数据不在我们服务器上，意味着换机不会自动跟过去，搬家是自己动手的操作；云同步在路线图上，做的时候会以端到端加密为前提。现阶段最稳的习惯是隔一阵导一个搬家包，存进自己信任的网盘。 ### TA 为什么会想起你？AI 伴侣主动消息的触发机制说明书 URL: https://foreverse.app/zh/blog/proactive-messages-mechanism Published: 2026-07-18 · Author: Deng Binjie AI 伴侣主动发消息是定时群发吗？在 Foreverse 里不是：每类触发都有明确条件——早安 8 点左右、晚安 10 点左右、想你了要超过一天没聊、低电量要低于 20% 且没充电、恶劣天气要你先手填城市。全局约束：默认关闭、免打扰 23:00–7:00、每天最多 2 条、手动关过的子项绝不复活、关掉无补发。这篇按开关逐个写清行为规格，包括 TA 什么时候不发。 Q: 主动消息收费吗？ A: 每条消息由后台真实生成一次，消耗一次文本调用：BYOK 走你自己的 Key 按供应商牌价，官方渠道按量扣积分。不触发就不生成、不花钱；每天最多 2 条的上限同时也是费用上限。没有配置可用文本模型时发不出，开启开关时会提示去配置。 Q: 恶劣天气关心开了却从来没响过，为什么？ A: 先查两件事：一是要在「TA 能感知的情境」里手填天气城市，没填城市这个触发永远不会满足——App 不申请定位权限，城市只能你自己给；二是触发条件是雨雪雷或极端温度，你的城市最近天气平稳就不会发。满足时每天也最多一次。 Q: 关掉总开关会漏掉纪念日祝福吗？ A: 会。总开关关闭时七类触发全部停止，纪念日也不例外，重新打开也不会补发错过的消息。想少打扰又不想漏大日子，可以留着总开关，只把早安、晚安、想你了这些高频子项关掉，保留纪念日祝福一项。 Q: 拒绝了通知权限还能用主动消息吗？ A: 能。Android 13 起发通知需要系统权限，开启总开关时会弹申请；拒绝后功能照常运转，TA 的消息正常落进你们的会话，只是不再有系统通知横幅，下次打开 App 才看到。想恢复横幅去系统设置给回权限即可。 ### 伴侣记忆管理实操：看、改、删，TA 记错了当场纠正 URL: https://foreverse.app/zh/blog/companion-memory-hands-on Published: 2026-07-18 · Author: Deng Binjie AI 伴侣记错事，在对话里纠正没用——对话会滑出窗口，记忆条目不会。这篇是实操手册：入口在伴侣聊天页右上 ⋮ 菜单的「记忆」，列表逐条可搜；铅笔改（单条上限 200 字）、垃圾桶删（确认弹窗显示原文）、开关停用（保留但不再进对话）、右上加号替 TA 记一条。改动从下一次对话生效；数据是手机本机文件，搬家包整包带走。 Q: 改了记忆，多久生效？ A: 从下一次对话开始生效。已经发出的历史消息不会被回改；正在进行的这一轮回复用的还是旧记录，下一轮起按新的来。验证方式很直接：随口聊一句相关话题，看 TA 接的对不对。 Q: 停用和删除有什么区别？ A: 停用是中间态：条目保留在列表里，但不再进入对话，开关随时拨回；删除是彻底移除，确认弹窗会把这条内容完整显示一遍，删了不可恢复。拿不准的先停用，确定不想留的再删。 Q: 「TA 记的」和「你记的」待遇一样吗？ A: 一样。两类条目都逐条进入对话，都能编辑、停用、删除，区别只是来源标签和记录日期。手动条目同样受单条 200 字限制，一条只记一件事效果最好。记忆提取随聊天自动进行，不额外计费。 Q: 记忆会上传到服务器吗？ A: 不会。记忆与人设、日记、纪念日同为设备本机的文件；换手机用搬家包整包导出再导入，TA 接着上次的记；删除伴侣时这些数据一并删除。BYOK 模式下对话请求直连你选的模型供应商，不经过我们的服务器。 ### 成人向角色卡的边界：五家模型供应商政策对照（附核实日期），与我们的态度 URL: https://foreverse.app/zh/blog/nsfw-cards-provider-policies Published: 2026-07-18 · Author: Deng Binjie 「哪家模型能聊成人向」是搜索量真实存在、但全网几乎没有诚实答案的问题。这篇不给承诺，给事实：OpenAI、Anthropic、Google、xAI、DeepSeek 五家现行条款对成人内容的写法逐家对照，每条标核实日期；讲清平台审核、供应商审核、模型拒答是三回事；最后亮我们自己的三条线。 Q: 成人向的卡用哪家模型不会被拒？ A: 没有「随便聊」的家，只有条款宽严的差异，而且政策随时会改。方法论比结论可靠：五家现行条款的对照表在正文里（每行标核实日期），按你的题材逐家读原文，挑政策相容的那家，并做好条款更新后重新评估的准备。任何给你打包票的答案，都没把「政策会变」算进去。 Q: API 会不会因为聊了擦边内容封我号？ A: 有真实可能。各家都在服务端跑检测系统：OpenAI 声明监控并执行政策，Google 明确对 API 滥用做自动扫描并保留数据 55 天用于违规检测，Anthropic 的 Safeguards 团队做检测与监控，处置从限流到终止账号（均于 2026-07-18 核实原文）。拒答只是模型层的软反馈，账号处置是供应商层的硬后果，两者别混为一谈。 Q: 平台审核和模型审核是一回事吗？ A: 是三回事。模型拒答是训练对齐的结果，换个写法可能就不拒了，它不是执法；供应商审核在服务端，扫描请求与输出、依据使用政策处置账号，这才是封号的来源；平台审核是你用的 App 或社区自己加的闸，比如上架商店的内容规则、社区分级。三层各管各的，哪一层出手，后果完全不同。 Q: 用 Foreverse，App 会替我过滤对话内容吗？ A: BYOK 对话不过滤：卡的导入不做内容审查，你的请求从设备直连你选的供应商，实际边界由那家的政策和模型行为决定。要过滤的是另一头——社区分享出来的内容有 nsfw 分级，18+ 的卡默认不展示，看之前要过年龄确认和两步确认。私下创作和公开分享，我们分开管。 Q: 有没有内容是无论哪家、无论怎么设置都不该碰的？ A: 有，而且各家条款在这一点上措辞完全一致：任何涉及未成年人的性内容是绝对红线，包括虚构和扮演场景。Anthropic 和 OpenAI 都写明会向 NCMEC（美国失踪与受虐儿童中心）报告，xAI 同样声明报告义务。我们的内容政策把它列为第一条硬线，发现即上报平台安全渠道，没有讨论空间。 ### 「破限」到底是什么？一篇给外行的诚实科普：词源、机制与各家政策 URL: https://foreverse.app/zh/blog/jailbreak-culture-honest-take Published: 2026-07-18 · Author: Deng Binjie 刷角色扮演社区总会撞上这个词：破限、破甲、发牌子。这篇给外行把它讲明白——词从哪来（2022 年 12 月 13 日 Reddit 的 DAN 帖）、技术上发生了什么、为什么时灵时不灵、各家供应商条款怎么写（附核实日期）、账号风险归谁。不教任何方法，只把机制和风险摆上桌，最后给一条我们认为更划算的路。 Q: 破限到底是什么意思？ A: 社区黑话，指用特殊提示词绕开 AI 模型的内容安全策略，让模型输出平时会拒绝的内容。同族黑话还有「破甲」（甲指模型的安全防御）、「发牌子」（触发平台的安全提示卡片）。它对应英文社区的 jailbreak，词根借自 iPhone 越狱。 Q: 破限会不会封号？ A: 风险真实存在，且写在条款里。OpenAI 使用政策明确把「绕过我们的防护措施」列为禁止项，违反可能失去访问权限；Anthropic 使用政策把「故意绕过产品内的能力限制或防护栏」列入平台滥用条目，处置手段包括限流、暂停和终止账号；Google 的生成式 AI 禁用政策也把「规避滥用保护或安全过滤器」列为禁止（三家条款均于 2026-07-18 核实原文）。执行松紧各家不同，但风险归属没有歧义：账号是你的，责任也是你的。 Q: 为什么有的角色卡写着「必须配破限才能玩」？ A: 通常是卡的题材在作者调试用的那家模型上容易触发拒答，作者便把某个引导性预设当成这张卡的运行环境标配。这句话其实在告诉你：这张卡的题材和默认模型的政策不匹配。你有两个选择：换一家政策与题材匹配的模型，或者接受对抗路线连带的全部风险。我们建议前者。 Q: 破限和预设是什么关系？ A: 载体和用途的关系。预设是打包的提示词结构加采样参数，本身是中性的配置工具，管文风、节奏、扮演纪律的预设占大多数；社区说的破限预设，是把绕过安全策略的引导语写进了预设里的那一类。工具同一个，用途分了叉。预设机制本身怎么分层、怎么遮蔽，我们有一篇专文讲。 Q: 用 BYOK 应用，App 会不会替我挡内容？ A: 以 Foreverse 为例：导入不做内容过滤，你的对话从设备直连你选的供应商，App 不在中间加一道自己的闸。实际能聊什么，由 Key 背后那家供应商的政策和模型行为决定。这也意味着责任链条很干净：协议是你和供应商直接签的，条款约束你，选择权也在你。 ### 酒馆预设是什么？分层、遮蔽关系与挑选思路一篇讲明白 URL: https://foreverse.app/zh/blog/tavern-presets-explained Published: 2026-07-18 · Author: Deng Binjie 导入一个预设，角色卡会不会被顶掉？为什么全局配了 A、这张卡却在用 B？这篇把预设讲成人话：它是一份「怎么跟模型说话」的打包配置，采样参数加提示词结构；四层生效顺序是会话、角色、全局、内置默认，越贴近这局越优先；最后给挑预设的三个问题和四个社区大部头的实数。 Q: 预设到底是什么东西？ A: 一份打包好的「怎么跟模型说话」配置文件：一半是采样参数（温度、输出长度这些旋钮），一半是提示词结构（系统指令怎么写、各模块按什么顺序拼进请求）。它管说话的方式，不管说话的人——人设、开场白、示例对话在角色卡里，预设不碰这些。 Q: 导入的预设会把我的角色卡顶掉吗？ A: 不会。预设和角色卡管的是两摊事：卡的人设内容照常注入，预设接管的是包裹这些内容的指令层和参数。真正会被顶掉的是同层的旧预设，以及你在全局设置里手写的系统提示词覆盖——预设整体接管系统提示位，两者不叠加，这是新手最容易误会的一处。 Q: 为什么我在全局选了预设，这张卡却没用上？ A: 生效顺序是四层：本次会话、当前角色、全局、内置默认，越贴近这局越优先，高层配了东西就盖住低层。这张卡大概率单独配过角色层预设，或者本次聊天应用过会话预设。想查就从会话层往下翻：本次聊天资料、当前角色设置、酒馆全局设置，哪层非空哪层说了算。 Q: 预设从哪找，新手该挑多大的？ A: 三个来源：中文社区流通的成套预设（QQ 群、论坛、网盘居多）、应用内社区页可直接逛和导入、以及桌面酒馆玩家导出的 JSON 文件。新手建议从小的开始：几十条以内、说明写得清楚的优先。上来就上两百多条的大部头，出了问题你分不清是卡的事、模型的事还是预设里哪一条的事。 Q: 预设在不同模型之间通用吗？ A: 格式通用，效果不通用。预设是标准 JSON，哪家模型都能导；但里面的指令是作者对着特定模型调出来的，换一家脾气不同的模型，防 AI 味的条目可能失灵，文风指令可能过火。社区预设一般会标注适配模型，跨模型用就当重新试一遍，别默认效果平移。 ### 酒馆群聊怎么玩才好玩？搭班子指南：请谁进群、怎么配戏才不冷场 URL: https://foreverse.app/zh/blog/group-chat-casting-guide Published: 2026-07-18 · Author: Deng Binjie 四张喜欢的卡拉进同一个群，结果一个人刷屏、三个人装雕塑——问题多半出在选角。这篇从运营视角讲群聊：放几个角色合适、四个功能位怎么配、话痨值怎么用、一个全员高冷的失败班底为什么必然冷场，以及每位发言背后的成本账。 Q: 群聊放几个角色合适？ A: 三到四个是甜区。两个人是对手戏不是群像；五个往上每轮发言机会摊薄，性格再鲜明也轮不到开口，token 成本还按人头涨。想要大场面，可以多拉几个进群但暂时禁言一部分，需要谁出场再解禁——在场感和成本都能兼顾。 Q: 为什么我的群聊老是一个人刷屏？ A: 最常见的是三件事叠加：那个角色的话痨值（talkativeness）比其他人高一截，加权抽选回回是它；你的消息老点它的名字，点名优先直接把话递过去；它的回复又爱提别人的名字，点名接力继续绕回它的话题圈。修法对应着来：调低它的话痨值、开场多点别人的名、或干脆切手动指定策略当几轮导演。 Q: 高冷寡言的角色就不适合进群吗？ A: 适合，但不能全是。寡言角色的话痨值低，加权抽选下开口少，恰好符合人设——问题只出在全员寡言的班底：没人主动接戏，每轮都靠顺序兜底硬轮，出来的戏像被按头发言。留一两个高话痨的角色当发动机，冷角色的偶尔开口反而更有分量。 Q: 群聊比单聊烧钱吗？ A: 按发言算，每位角色开口一次就是一次模型调用，一轮引出两三个人接话就是两三次调用。另一头是角色卡注入方式：单独注入只带当前发言者的卡最省；合并全员把所有成员的卡都塞进每次请求，四张卡就是四份常驻 token。长剧情用单独注入、关键对手戏切合并，是比较省的开法。 Q: 在手机上怎么建群？ A: 角色页长按任意一张卡进入多选，勾满两张以上可开聊的卡，底部会出现建群条，点「新建群聊」即成。发言策略、成员启停、自动模式都在聊天页右上 ⋮ 菜单的「群聊设置」里，随时可改，改了立刻生效。 ### Quick Reply 自动化玩法：把每次都要敲的那句话做成一颗按钮 URL: https://foreverse.app/zh/blog/quick-reply-automation-guide Published: 2026-07-18 · Author: Deng Binjie 酒馆的 Quick Reply（自动按钮）是把重复劳动钉成按钮的机关：一键剧情总结、换景模板、无声导演批注、AI 回复后自动计数，四个可直接照抄的配方。按钮内容以 / 开头就是 STscript，手机端跑的是桌面同源的命令子集，不认识的命令安全跳过；四档自动化模式加死循环检测兜底，脚本里只有触发生成的步骤才花钱。 Q: 酒馆的 Quick Reply 是干嘛用的？ A: 把你反复要做的操作钉成输入框上方的一颗按钮。按钮内容是纯文本时，点一下就替你发出那句话（或填进输入框）；以 / 开头时就是一段 STscript 脚本，能做条件、循环、变量这类复杂机关。按钮还能挂自动时机，比如每次 AI 回复后自动执行，从「快捷键」升级成「自动化」。 Q: 有没有办法一键让 AI 继续、总结、换视角？ A: 有，这正是自动按钮的主场。总结：建一颗纯文本按钮，内容就是你每次都要敲的那句「暂停剧情，用三句话总结当前进展」，点一下即发。继续：脚本按钮写 /continue。换视角、换景：用 /setinput 把模板填进输入框，改两个词再发，比全文重敲快得多。三颗按钮各占一个位置，装完就是你的导演台。 Q: STscript 在手机上能跑吗？ A: 跑的是桌面同源的命令子集：变量（/setvar、/getvar、/incvar 等）、流程（/if、/while、/times）、发送与生成（/send、/sendas、/trigger、/continue）、输入框与弹窗（/setinput、/input、/buttons）、候选与群成员管理都在。输入框打一个 / 会弹出命令菜单点选补全；遇到暂不支持的命令会安全跳过，不执行也不动你的数据。完整清单在 设置 → 角色卡与酒馆 → 酒馆命令帮助。 Q: 自动化会不会失控刷屏、偷偷花钱？ A: 有三道闸。模式闸：自动化分四档（酒馆兼容/发送前确认/只填输入框/禁用），不放心就切「发送前确认」，每次自动发送前都先问你。循环闸：检测到疑似自动循环会自动把模式降到「发送前确认」并提示。账目闸：脚本本身不花钱，只有 /trigger、/continue 这类触发生成的步骤按正常对话计费；运行记录里能查每一步做了什么、有没有触发生成。 ### 酒馆正则入门：它改的是显示还是历史？第一次用先分清这个 URL: https://foreverse.app/zh/blog/regex-scripts-for-beginners Published: 2026-07-18 · Author: Deng Binjie 酒馆正则就是自动跑的查找替换：藏思考链、去 OOC 旁注、把状态栏 marker 换成美化界面，全是它。但第一次用之前要分清三档执行时机——「双向」在消息写进记录前就替换、改动是永久的；「仅显示」只改你看到的；「仅提示词」只改模型看到的。这篇玩家向教程带四步上手：导入带正则的卡、认时机、从模板建第一条、出问题单条排查。 Q: 酒馆的正则脚本是干嘛的？ A: 自动执行的查找替换规则。每条规则盯着一类文本（用户输入、AI 回复、世界书等，可多选），命中就按替换串改写。玩家侧最常见的三个用途：把回复里的思考段折叠掉、去掉 (OOC:…) 这类旁注、把模型输出的纯文本 marker 替换成状态栏或主题化界面——美化包的核心机制就是最后这条。 Q: 导入的卡带正则，会不会乱改我的聊天记录？ A: 美化类正则不会：它们走「仅显示」时机，只在渲染时替换，聊天记录文件里存的还是模型的原话，HTML 模板也不占 token。会动记录的是「双向」时机的规则——替换发生在消息写盘之前，改动是永久的（比如把思考标签从记录里删干净就靠它）。在全局正则页里每条规则的时机都有标签，导入后扫一眼就知道谁会动记录。 Q: 正则美化手机上能用吗？和桌面酒馆一致吗？ A: 能用。Foreverse 的正则引擎按桌面酒馆的语义实现：作用位置（用户输入/AI 回复/世界书/Slash/思考链）、三档执行时机、$1 反向引用与 {{match}} 替换、按楼层深度限定生效范围，行为一致。桌面上调好的美化卡导入手机端直接生效，导入时还会自动打开该卡的富渲染开关。 Q: 正则写坏了会把聊天弄崩吗？ A: 不至于。最坏的情况是这条规则把显示文本替换成了空，渲染侧有兜底：清空时回退显示原文，不会白屏，也不会吞掉正文。修复路径是回全局正则页把它关掉或删掉；「仅显示」档的错误不碰记录文件，关掉规则一切如初。真正要小心的只有「双向」档——它写进记录的改动没有撤销键。 ### 世界书递归是什么？触发链、深度限制与一次失控案例 URL: https://foreverse.app/zh/blog/lorebook-recursion-explained Published: 2026-07-18 · Author: Deng Binjie 「为什么加了一条设定，好几条一起冒出来？」这是世界书递归扫描在干活：已命中词条的正文也被当作扫描源，词条可以召唤词条。这篇机制说明书讲清链条怎么走、在哪停（层数、预算、去重三道闸）、三个词条级开关各管什么，附一个五条连锁的可复现失控案例和两件自带的验证工具。 Q: 世界书条目会互相触发吗？ A: 会，前提是打开「递归扫描」（Foreverse 里默认是关的，开关在酒馆全局设置的「世界书与渲染」）。打开后，已命中词条的正文也参与关键词扫描，正文里点到谁的关键词，谁就跟着注入。同一词条一次生成至多激活一次，所以 A 提 B、B 又提 A 不会转圈。 Q: 世界书递归扫描是什么意思？ A: 指扫描源多了一层。普通触发只扫最近几轮对话找关键词；递归扫描把已命中词条的内容也当作扫描文本，让词条能链式召唤相关词条：聊到门派，门派档案里点名的佩剑也自动跟上。链条停在三个条件上：没有新命中、达到最大递归层数、token 预算耗尽。 Q: 为什么加了一条设定，好几条一起冒出来？ A: 大概率是新条目的正文里点了其它条目的关键词，或者被别的条目点名，递归把它们串成了一串。打开聊天页 ⋮ 菜单的「上次生成请求」看世界书明细，每条注入都标着命中关键词，链条一眼可见。不想被链子带出来的条目，在词条编辑器里勾「不参与递归扫描」。 Q: 最大递归层数设多少合适？ A: SillyTavern 官方文档的口径：设 1 等于关闭递归，2 允许链条走一跳，3 允许两跳，依此类推。日常两跳足够覆盖「门派带出佩剑、佩剑带出剑灵」这类两级脉络。层数是全局旋钮；个别词条的精细控制，交给「不参与递归扫描」「阻断后续递归」「仅递归时激活」三个词条级开关更划算。 ### 睡前听书方案：定时关闭、音色、缓存和不吵醒室友的细节 URL: https://foreverse.app/zh/blog/bedtime-listening-setup Published: 2026-07-18 · Author: Deng Binjie 睡前听小说的完整设置清单：睡眠定时 15/30/60/90 分钟四档怎么选、深夜听不吓人的音色和语速搭配、提前缓存整章防半夜断网、锁屏通知栏控制和不吵醒室友的细节，外加「听到哪明早从哪读」的进度接力。免费系统 TTS 就能起步，全部步骤按关灯前五分钟的顺序排好。 Q: 睡前听小说用什么 App，能定时吗？ A: 挑应用盯三样：有睡眠定时、能提前缓存、进度和阅读打通。以 Foreverse 为例，阅读器底部控制栏点「听」进播放器，睡眠定时给 15、30、60、90 分钟四档，到点自动停；整章可以在 Wi-Fi 下提前缓存；耳朵听到哪，阅读进度就记到哪，第二天翻开书接着读。系统 TTS 引擎完全免费，先用它把习惯养起来。 Q: 听书定时关闭一般设多久合适？ A: 从 30 分钟档起步。多数人入睡在这个窗口内，播完自动停，不会整夜耗电也不会把书「听」过去几十章。躺下就困的人用 15 分钟档；习惯听着入睡又浅眠的人别选 90 分钟，声音持续的时间越长，越容易在浅睡期把你重新拉醒。定时到点停在哪一段，进度就记在哪一段，第二天不用倒回去找。 Q: 晚上听书哪个音色不吓人？ A: 两条经验：语速放慢，选温暖或低沉的声线。白天通勤用 1.5 倍速赶进度没问题，睡前建议降到 0.75 到 1 倍，语速本身就是刺激源。音色上避开冷峻快节奏的悬疑腔，言情种田配温暖女声、历史修仙配低沉男声的搭配深夜同样成立。系统 TTS 的默认音色偏平，反而适合催眠；在线 AI 音色更有起伏，音量记得比白天再调低一档。 Q: 半夜网络断了，听书会停吗？ A: 分引擎。系统 TTS 装了离线语音包就完全不依赖网络，飞行模式也能听。在线 AI 音色按分段合成，睡前在 Wi-Fi 下点「下载整章」把接下来的内容缓存好，已缓存的段落播放不需要网络；没缓存的段落会等有网时排队补合成。缓存还管钱包：合成结果按章节、音色、模型落盘，重听不重复扣费。 ### token 是什么？给只想看小说的人算一笔账 URL: https://foreverse.app/zh/blog/token-explained-for-readers Published: 2026-07-18 · Author: Deng Binjie token 是 AI 读写文字的计量单位，也是计费单位。我们拿一章 3065 字的真实网文实算：不同家的分词器数出 1800 到 4200 个 token 不等，约合一个汉字 0.6 到 1.4 个 token。这篇全程用「章、段、几分钱」当单位，把「续写一段为什么只要一分多钱」「缓存为什么能打到零头价」算到你能核对的粒度。 Q: AI 按 token 收费，到底是什么意思？ A: token 是模型读写文字的最小单位，直观理解成「字或词的碎片」：模型每读一个 token、每写一个 token 都记一次账，账单等于读写总量乘以单价。对读者的实际意义是，你付的钱跟「模型这次读了多少前文、写了多少新内容」成正比，跟你这本书总共多长无关，因为每次请求只送需要的部分进去。 Q: 一章 3000 字的小说大概是多少 token？ A: 看谁来数。我们拿一章 3065 字的真实网文实算：OpenAI 新版分词器数出约 2,800 个，旧版数出约 4,200 个；DeepSeek 官方给的换算口径是一个汉字约 0.6 个 token，即约 1,800 个。同一章书，各家能差出一倍多，所以别背精确数字，记住量级：一章几千个 token，实际扣费以每次请求返回的用量统计为准。 Q: 为什么说续写一段只要几分钱？ A: 一次续写，模型读进去的是相关前文加设定，一万个 token 上下，写出来一段三五百字。按 DeepSeek 2026 年 7 月牌价把这笔账换算掉：读的部分约一分钱，写的部分不到两厘，合计一分二厘左右。走 Foreverse 官方渠道的实测均值是 19 credits，约一分四厘。两个口径都指向同一个量级：一段续写是「分」级的开销。 Q: 有的模型「思考」也收费吗？ A: 收。推理型模型在给出正文前会先产出一段思考过程，这部分 token 按输出价计费，哪怕你在界面上看不到全文。同一段续写任务，思考型档位的实际消耗可能是普通档位的数倍，这是「单价看着便宜、账单却更贵」的常见原因。给这类模型设输出上限时也要把思考预算算进去，否则正文会被截断。 Q: 怎么查自己每次续写用了多少 token？ A: 不用信任何文章的估算，包括这篇。每次 API 调用的返回里都带用量统计字段，写着这次读了多少、写了多少、命中缓存多少。Foreverse 的「API 请求记录」（设置 → AI 模型与服务）把这些字段逐条摆出来，连续续写两次，对比第二次的缓存命中数，你会亲眼看到本文最后一节说的折扣。 ### 2026 年免费用 AI 续写小说的四条路，各自的天花板实测 URL: https://foreverse.app/zh/blog/free-models-for-continuation-2026 Published: 2026-07-18 · Author: Deng Binjie 2026 年零成本用 AI 续写小说的四条路逐条核实：注册送 5000 积分约 260 段续写；智谱 GLM-4.7-Flash 官方免费调用、Gemini AI Studio 免费层限次不限期；硅基流动实名领 16 元代金券可试旗舰；本地 Ollama 零 API 成本。每条路的真实天花板（额度、限速、机能、文笔）如实标注，附组合走法。 Q: 有没有完全免费的 AI 续写？ A: 有，而且不止一条路。长期免费的是 BYOK 免费档：智谱 GLM-4.7-Flash 在开放平台官方标注免费调用，Google AI Studio 的 Flash 档免费层限次数不限期限（均为 2026 年 7 月核实）。一次性免费的有 Foreverse 注册送的 5000 积分（约 260 段）和各家新号体验金。完全离线的是本地 Ollama 跑开源模型。四条路都不要钱，差别在天花板的位置。 Q: 学生党没钱，想用 AI 写文，推荐哪条组合？ A: 日常推进接智谱 GLM-4.7-Flash（免费、注册 bigmodel.cn 拿 Key 即用），写到在意的章节切 Foreverse 注册送的 5000 积分用官方渠道的旗舰模型，两边随时并存。这套组合零投入，靠免费档扛量、送的额度保关键章节质量。硅基流动实名认证的 16 元代金券留着试模型：想知道某个付费旗舰值不值得充，先拿代金券写几段看效果。 Q: AI 写小说不充钱能用多久？ A: 分路径。BYOK 免费档没有总量截止日：智谱免费模型和 AI Studio 免费层都是长期政策（后者每天限次、太平洋时间午夜重置），只要政策不变就一直能写。一次性额度会到头：注册送的 5000 积分约 260 段、8 到 13 万字。本地 Ollama 理论上无限，电费不计。所以「用多久」的答案取决于你接受哪种天花板：限速、限量，还是限机能。 Q: 免费模型写出来的质量比付费的差多少？ A: 差距真实存在，但位置分场景。免费档模型（GLM-4.7-Flash、Gemini Flash 免费层）写日常推进、对话、过渡段够用；长剧情的伏笔管理和文风贴合度不如旗舰，这是我们在两场双盲横评里反复看到的档位差。务实的用法不是二选一：免费档跑量、旗舰保关键章节，在支持逐段换模型的阅读器里两档并用没有切换成本。 ### Grok 能写小说吗？复读实测、适用边界与什么时候别用它 URL: https://foreverse.app/zh/blog/grok-continue-novel-howto Published: 2026-07-18 · Author: Deng Binjie Grok 4.5 写小说的直答页：360 轮双盲横评里它是唯一在两种文体都掉进复读循环的模型，宫斗场评审一致垫底、二十轮剧情停在案发当日下午。这篇讲清复读为什么不是你的问题、Grok 真正能打的场景、什么时候别用它，附 2026 年 7 月核实的 xAI 牌价与国内接入现状。 Q: grok 写小说怎么样，有人测过吗？ A: 测过。我们让九个大模型在两本中文书（传统玄幻与宫斗古言）上各连续续写 20 轮、累计 360 轮，双盲评审排名。Grok 4.5 在宫斗场被两个互不知情的评审一致排在末位：后段叙述与台词逐字复读，二十轮剧情停在案发当日下午；玄幻场中段也出现同两段文字逐字循环三遍。短程生成可用，长跑连续续写是它实测的雷区。 Q: grok api 国内能用吗？ A: 主要门槛在支付而不在技术：xAI 的充值不支持支付宝、微信和国内银行卡，只收海外信用卡（2026 年 7 月核实）；国内网络直连 api.x.ai 的稳定性也因网络环境而异。社区里常见的做法是经聚合中转平台间接接入，Foreverse 的 BYOK 预置目录里有整组「聚合与中转」供应商，多花一点差价换省事。 Q: grok 写文老是重复，是我的问题吗？ A: 不是。同一套提示、同一个实验协议下，多数模型没有这种高频发作，而 Grok 4.5 在玄幻和宫斗两种文体上都发作了，这是模型级的失效，不是提示词能根治的。你能做的是缓解：发现开始复读时换一个模型续下一段，或者把复读段删掉重写，循环就被打断了；以及别把几十段的长链一口气交给它。 Q: Grok 便宜的 fast 档能拿来写文吗？ A: 按 xAI 2026 年 7 月牌价，grok-4.1-fast 每百万 token 输入 0.2 美元、输出 0.5 美元，比旗舰低一个量级。但我们的横评只测了 grok-4.5，fast 档没有续写质量数据，这里不替它背书。想试的话用短程任务自己盲测：同一段开头各续几段，跟你常用的模型对比着读，再决定要不要把日常草稿交给它。 ### 智谱 GLM 续写小说：唯一双文体都进前三的模型，和它的半角引号顽疾 URL: https://foreverse.app/zh/blog/glm-continue-novel-howto Published: 2026-07-18 · Author: Deng Binjie 智谱 GLM 5.2 是我们两场双盲横评（890 万字传统玄幻 + 《甄嬛传》宫斗古言）里唯一双榜都进前三的模型，代价是跨文体复现的半角引号顽疾：玄幻场 20 轮 108 处，宫斗场原样复发。这篇按一次完整开书流程走：bigmodel.cn 实名建 Key、免费档 GLM-4.7-Flash 试笔、旗舰 5.2 正式续写、撞上引号后的三层绕行，附 2026 年 7 月核实牌价。 Q: 智谱清言写小说怎么样？ A: 智谱清言是面向普通用户的聊天助手，模型同源但容器不同：聊天框贴不下一本书，会话之间前情也不互通。想接着一本书往下写，走 API 路线更顺：在 bigmodel.cn 建 Key，填进 BYOK 阅读器。模型能力本身有名次可查：GLM 5.2 在我们的双盲横评里玄幻场第 3、宫斗场 2-3，唯一双文体都进前三。 Q: GLM 续写网文行不行？ A: 行，有实测：玄幻场（六模型双盲）它排第 3，评语「内容层贴近（战术商议、单字传讯都对味）」；宫斗场评语「限知观察质感最纯」。要带一个预期进场：它习惯用半角引号写对话，中文书里观感出戏，缓解靠提示词约束、对话密集章节逐段换模型，或对已生成段落编辑重生。 Q: 国产模型里哪个写文最稳？ A: 「稳」有两种读法。单场峰值看 DeepSeek 双档：玄幻第一是 V4 Flash，宫斗第一梯队是 V4 Pro；跨文体稳定性看 GLM 5.2，唯一双榜都进前三的模型。读单一文体的书，按那张榜的第一名挑最划算；文体杂、或者书还没定型，GLM 是更保险的开局牌。 Q: GLM 免费档和付费档差多少？ A: 免费档 GLM-4.7-Flash 官方长期免费、200K 上下文，推荐场景明确列了中文写作与角色扮演，日常推进和过渡段够用。但进两张榜前三的是付费旗舰 5.2，免费档没有盲评名次，两者的口碑别混着用。务实组合：免费档跑量，5.2 保关键章节，同一把 Key 里切换。 ### Kimi 续写小说实测：长上下文是真优势还是伪需求 URL: https://foreverse.app/zh/blog/kimi-continue-novel-howto Published: 2026-07-18 · Author: Deng Binjie Kimi K2.6 的 256K 上下文是真能力，「窗口大就续得像」是伪推论：九模型《甄嬛传》双盲横评它排 5-7 名，评语「越写越收敛」与窗口无关；4k 到 20 万 token 的五档对照实验证明，整本塞入不加指令照样比喻密度 10 倍于原著，且每段按全书输入计费。附 2026 年 7 月官方牌价（输入 ¥6.5/百万 token）、接入步骤与长上下文真正适用的场景清单。 Q: Kimi 可以续写小说吗，效果如何？ A: 可以，有名次可查：九模型《甄嬛传》双盲横评里 Kimi K2.6 排 5-7 名，评语是「唯一显著逆向改善的系统：开局是网文暴怒腔，越写越收敛回宫斗正轨」；同一协议的玄幻链上它有轻症复读，整段照抄自己前几轮的产物。总体中游偏稳，长跑撞见似曾相识的整段就重摇一次或换个模型打断。 Q: Kimi 的长上下文对写文有用吗？ A: 分任务。整卷总结、全书人物梳理、世界书起草这类要一口气读完全书的活，256K 窗口是真优势；反复续写不是：我们的五档上下文对照实验里，把 27 万字原著整本塞入、不加指令，产出照样满篇套话比喻，密度是原著的 10 倍，而且每段续写都按全书输入计费。续写要的是按相关性挑出的那一小撮上下文。 Q: 月之暗面的 API 多少钱、怎么接？ A: K2.6 官方牌价（2026-07-18 核实）：输入未命中缓存 ¥6.5、命中 ¥1.1、输出 ¥27，单位每百万 token，预付费制，余额为零请求返回 402。在 platform.moonshot.cn 注册建 Key 并充值，填进 BYOK 阅读器（Foreverse 预置 Moonshot 条目，host 是 api.moonshot.cn/v1），用页内实测功能验证连通后，导入书长按「续这里」即可。 Q: Kimi 网页版免费，和 API 是一回事吗？ A: 两个账户体系。网页与 App 的 Kimi 助手面向普通用户，基础聊天免费（高峰时段可能限流，政策以官方页面为准）；API 在开放平台单独注册、预充值、按 token 计费，两边余额互不相通。把 Kimi 接进阅读器续写一本书，走的是 API 这条路，网页版的免费额度帮不上。 ### Gemini 续写小说实测：免费档能白嫖到什么程度，短板在哪 URL: https://foreverse.app/zh/blog/gemini-continue-novel-howto Published: 2026-07-18 · Author: Deng Binjie AI Studio 免费档 2026 年 7 月核实：Flash 级每分钟约 10-15 次、每天数百到一千多次请求，长期有效不用绑卡；但免费层只覆盖 Flash 级，我们双盲横评排名的是付费层的 Gemini 3.1 Pro——玄幻场曾因一句标注措辞 20 轮全部重写开场，措辞修复后甄嬛传场回到 4-6 名。这篇把「免费」和「能打」两笔账分开算，附拿 Key 与接进手机阅读器的步骤。 Q: Gemini 免费的 API 可以拿来写小说吗？ A: 可以，两个边界内：免费档只覆盖 Flash 级模型（2026 年 4 月起 Pro 级归付费层），配额大约每分钟 10-15 次、每天数百到一千多次请求，按 Google Cloud 项目计而不是按 Key 计。一段续写等于一次请求，正常阅读节奏用不完每天的量，连续快速重摇才会撞每分钟限次。另外官方条款写明免费层的请求与产出可能被用于改进产品，喂未发表的手稿前值得停一秒。 Q: AI Studio 的 Key 怎么弄？ A: 用 Google 账号登录 aistudio.google.com，进 API keys 页点 Create API key，选一个新建或已有的 Google Cloud 项目，复制 AIza 开头的字符串，全程不用绑卡。和 OpenAI、Anthropic 不同，AI Studio 之后还能回平台查看 Key 明文。拿到后填进支持 BYOK 的阅读器（如 Foreverse 的模型供应商页），用页内实测功能点一次文本聊天验证连通。 Q: Gemini 写长篇会不会越写越差、老是从头重写？ A: 我们实测撞到的是急性病「从头重写」：Gemini 3.1 Pro 在玄幻续写里 20 轮全部跳回原著末尾重写开场，根因是上下文里一句「仅供参考」式的标注说明被它读成「这些段落不算正文」。把说明改成两句显式声明（已续写内容是已发生的正文、必须从最后一段继续；文风以原著为基准）后，两本书合计 28 轮零复发。手工喂上下文时写清这两句，就能避开这个坑。 Q: 免费档的 Gemini 和横评里排名的是同一个模型吗？ A: 不是。我们双盲横评排名的是付费层的 Gemini 3.1 Pro（《甄嬛传》场 4-6 名）；免费档只有 Flash 级模型，我们没有它的盲评名次，也不硬编。免费档适合当零成本试镜场：拿自己那本书的固定一段让它连续续写几轮，打乱顺序盲读，再决定要不要为付费档或别家模型花钱。 Q: 中国大陆能用 AI Studio 的免费档吗？ A: AI Studio 与 Gemini API 的官方可用地区名单（截至 2026-07）不含中国大陆，直连走不通，网络环境需自理且合规性自行判断。省事的替代有两条：Foreverse 官方渠道目录里有 Gemini 条目，登录按积分计费，注册送 5000 credits；或改用国内可直连的免费档，比如智谱 GLM-4.7-Flash。 ### Claude 怎么续写小说？从网页版长度天花板到 API 接进手机阅读器 URL: https://foreverse.app/zh/blog/claude-continue-novel-howto Published: 2026-07-18 · Author: Deng Binjie 想用 Claude 续写手机里的小说：claude.ai 的 Projects 大书自动切检索模式、写出的段落散在会话里，续写要走 API。这篇给四步教程：console.anthropic.com 开户建 Key、填进 BYOK 阅读器、设默认模型、第一段续写落分支；附 2026 年 7 月官方牌价折算（Opus 一段约五毛钱）与双盲横评里 Claude 的真实名次。 Q: Claude 能接着我手机里的小说继续写吗？ A: 能，走 API 路线：在 console.anthropic.com 开户建 Key，填进支持 Anthropic 协议的 BYOK 阅读器（比如 Foreverse），导入 txt 或 epub 后长按选段点「续这里」。整本书的上下文由阅读器按相关性组装，不用每次手工喂前文；新内容落在分支上，原文一个字不动。 Q: 我充了 claude.ai 的 Pro，还要再花钱吗？ A: 要。Pro 订阅买的是网页与 App 的聊天用量，不含任何 API 额度；API 在 console.anthropic.com 单独开户、预充值、按 token 计费，同一个邮箱登两边但账单互不相通。按 2026 年 7 月牌价折算，Sonnet 4.6 一段续写约三毛人民币，Opus 4.8 约五毛。 Q: Claude 写中文小说效果怎么样，值得充吗？ A: 我们两场双盲横评里它都排第 4-5 名：对话机锋全场最锋利，但尾段半角标点混排、波动大，两场第一梯队都是别家。想要 AI 有自己笔感的场景（同人重构、机锋密集的对话戏）值得；追求「像原著续章」的无缝长跑，玄幻场第一是 DeepSeek V4 Flash、宫斗场第一梯队是 V4 Pro 和 GPT-5.6 Terra，都比它便宜得多。 Q: 中国大陆能用 Claude 续写小说吗？ A: Anthropic 官方支持地区名单（截至 2026-07）不含中国大陆，直接开官方 Key 走不通。合规的路有两条：用 Foreverse 官方渠道里的 Claude 条目（Sonnet 4.6 与 Opus 4.8），登录后按积分计费；或者接入你自己信得过的 Anthropic 兼容第三方端点，BYOK 支持自定义端点，但中转服务的合规性与稳定性要自行判断。 ### 我们把角色卡创作 Skill、47 份分类指南和 AI 味检测器全部开源了 URL: https://foreverse.app/zh/blog/open-sourcing-character-card-skills Published: 2026-07-17 · Author: Deng Binjie character-card-skills 仓库上线：两个 agent skill（写卡 + 聊后调优）、47 份分类创作指南、15 张过检测的原创卡（酒馆可直接导入的 v2 JSON + v3 PNG 嵌卡都给了），以及那个在 LLM 评委 gold 正确率只有 12% 之后、按真实读者判断校准出来的规则版 AI 味检测器。代码 MIT，内容 CC BY 4.0。Claude Code、Cursor、Codex 都能跑，在 Foreverse 安卓端是内置的。 Q: character-card-skills 仓库里到底有什么？ A: 两个 agent skill（character-card-author 写卡 + chat-quality-doctor 聊后调优）、47 份分类创作指南、8 份分模型 RP 症状档案、15 张原创卡（12 中 3 英，每张给 card.md 源文件、chara_card_v2 JSON 和 v3 PNG 嵌卡三种形态）、按真实读者判断校准的 AI 味规则检测器和它的 gold 回归集、一个零依赖的转卡脚本。代码 MIT，卡和文档 CC BY 4.0。 Q: 卡能直接导入酒馆（SillyTavern）吗？ A: 能。每张卡都预转换好了：v3 PNG 拖进酒馆的角色导入就能用，偏好 JSON 的前端用旁边的 v2 JSON。世界书、备选开场白、示例对话都嵌在卡里。我们在 SillyTavern 和 Foreverse 里各导入验证过一遍。 Q: skill 在哪些工具里能跑？ A: 所有认 SKILL.md 开放标准的 agent：Claude Code、Cursor、Codex、Gemini CLI，装法就是拷一个文件夹。在 Foreverse 安卓端这套 skill 是内置的——书架 agent 写完卡还能顺手评分、画封面、直接入库。 Q: 为什么用规则检测器而不是让大模型来评「像不像 AI」？ A: 因为两个我们都测过。双盲实验里 LLM 评委彼此 86% 意见一致，在人类共识锚点上的正确率却只有 12%——一致地把真人写的卡判成 AI。规则检测器按真实读者点名的病句校准，每张进仓的卡都要在 CI 里过它。它的诚实边界：结构完美、密度均匀的模板化 AI 文案，它也抓不到。 ### We Open-Sourced Our Character Card Authoring Skills, 47 Genre Playbooks and the AI-Flavor Detector URL: https://foreverse.app/blog/open-sourcing-character-card-skills Published: 2026-07-17 · Author: Deng Binjie character-card-skills is live: two agent skills (card authoring + chat-quality triage), 47 genre playbooks, 15 original cards shipped as SillyTavern-ready v2 JSON and v3 PNG, and the rule-based AI-flavor detector we calibrated after our LLM judges failed gold calibration at 12%. Code MIT, content CC BY 4.0. Runs in Claude Code, Cursor, Codex, Gemini CLI — and natively in Foreverse on Android. Q: What exactly is in the character-card-skills repo? A: Two agent skills (character-card-author and chat-quality-doctor), 47 genre-specific writing playbooks, 8 per-model roleplay symptom profiles, 15 original character cards (12 Chinese, 3 English) shipped as card.md source plus chara_card_v2 JSON plus v3 PNG-embedded cards, a rule-based AI-flavor detector with its gold regression set, and a zero-dependency md-to-card converter. Code is MIT; cards and docs are CC BY 4.0. Q: Do the cards work in SillyTavern? A: Yes. Every card ships pre-converted: a v3 PNG you can drag into SillyTavern's character import, and a v2 JSON for frontends that prefer it. Lorebooks, alternate greetings and example dialogues are embedded. We test imports in SillyTavern and in Foreverse. Q: Which coding agents can run the skills? A: Anything that reads the open SKILL.md convention: Claude Code, Cursor, Codex, Gemini CLI. Install is a folder copy. The same skills also run natively inside Foreverse on Android, where the agent can score, illustrate and import the card it just wrote. Q: Why trust a rule-based detector over an LLM judge? A: Because we measured both. In a double-blind panel our LLM judges agreed with each other 86% of the time and were right 12% of the time on human-consensus anchors — they systematically rated human-written cards as AI. The rule-based detector is calibrated against what real readers actually flagged, and it gates every card in the repo through CI. Its honest limit: perfectly templated AI copy still passes. ### Continue Sherlock Holmes With AI: The Pastiche Tradition Goes Personal URL: https://foreverse.app/blog/continue-sherlock-holmes-with-ai Published: 2026-07-17 · Author: Deng Binjie Sherlock Holmes has been continued by other hands since the 1890s — thousands of pastiches, an estate-endorsed novel, 250+ screen portrayals. Since January 1, 2023 the entire canon is public domain in the US. A practical guide: where to get the source text (Project Gutenberg), why Watson's voice is a natural style anchor, what long-run failures to expect from real 360-round data, and how branching handles the Reichenbach what-if. Q: Is Sherlock Holmes really public domain now? A: Yes. In the US, 50 of the 60 canonical stories had already aged into the public domain by 2013, and the Klinger v. Conan Doyle Estate litigation (decided 2013, affirmed by the Seventh Circuit in 2014) confirmed the Holmes and Watson characters were free to use. The last ten stories, collected in The Case-Book of Sherlock Holmes (1927), expired one by one until January 1, 2023, when the final two entered the US public domain. In the UK and other life-plus-70 countries, copyright ended decades earlier — Conan Doyle died in 1930. What remains off-limits is implying endorsement by the estate, which is a trademark matter, not copyright. Q: Will the AI actually sound like Conan Doyle? A: Honest answer: it can hold the register — formal Victorian narration, Watson's clinical warmth, period vocabulary — but surface mimicry is not the same as the real thing, and our own evals keep proving it. In 360 rounds of continuation testing we saw the best prose mimic in the field reproduce an author's comma rhythm almost perfectly, then rewind the entire plot in the final rounds. Expect a competent impersonation that needs your editing, not a resurrection. Q: Can I publish my Holmes pastiche? A: The copyright side is clear: the original canon is public domain, so continuation and publication are lawful in ways that fanfiction of in-copyright works is not — an entire commercial pastiche industry already exists. Two practical caveats: don't imply official endorsement (trademark territory), and if AI wrote part of it, disclose that wherever you publish — most platforms now have explicit AI-content rules and quietly breaking them gets works removed. Not legal advice. Q: Which model should I use for Victorian-era prose? A: We haven't run a Victorian-English arena yet, so we won't invent a ranking. What our Chinese-fiction arenas did establish transfers as a method: no ranking survives a genre change (our fast-plain-prose champion fell to sixth on an ornate court novel), so test two or three models on the same scene from your actual book and compare. In Foreverse you can switch models per scene with your own keys — 62 providers — so the test costs a few minutes, not a subscription. ### Story Continuation Prompts That Survive Long Runs: Field Notes and Four Copy-Ready Templates URL: https://foreverse.app/blog/story-continuation-prompting-guide Published: 2026-07-17 · Author: Deng Binjie Bare 'continue this' prompts decay measurably over consecutive rounds — we have the 20-round benchmark data. These field notes split the working prompt into an instruction layer and a material layer, explain the continuity directive that took one model from 20/20 restart loops to 0/20, and include four copy-ready English templates (fast action, interior-emotional, suspense, literary restraint), each with banned-phrase lists and a note on when to use it. Q: Can a prompt fully clone the original author's style? A: No — close, not perfect, and the gap is measurable. In our 20-round benchmark, no model matched the original's sentence-length burstiness: the source book's coefficient of variation was 0.84 and all six models landed between 0.46 and 0.59. Better context labeling reversed the decay direction in follow-up runs — one arm climbed from 0.56 back to 0.64 across rounds instead of sliding — but nothing we tried closed the gap fully. A good template plus source labeling plus the right model gets you to 'stays immersive on a straight read-through'; anyone claiming pixel-perfect style cloning from a prompt is overselling what prompts do. Q: How long should the excerpt I paste be? A: Enough to carry the voice, not the whole book. Style lives in the material layer, so a few thousand words of the most recent text — the run-up to your continuation point — outperforms a plot summary of everything. Our benchmark chains ran on a fixed 16k-token window (roughly 12,000 English words) and that was sufficient context for the best pairings to read like the original. In a reader app that assembles context automatically you can skip this question; pasting by hand into a chatbot, favor the last several chapters over scattered highlights. Q: Why do these templates ban phrases instead of describing the style? A: Because negative constraints are checkable and adjectives are not. A model cannot reliably act on 'write with more tension', but 'do not have the narrator announce danger before the character sees it' is a rule it can follow sentence by sentence — and that you can verify on read-through. The same logic applies to cliché lists: naming the exact constructions you keep seeing ('breath caught', 'time seemed to slow') removes them far more reliably than asking for 'fresh prose'. Q: Do these templates work in any AI tool? A: Yes — they are plain instructions with no app-specific syntax, so they paste into any chatbot or writing tool. Two adjustments travel with them: supply the material layer yourself (recent story text, plus a line stating that earlier AI continuations are canon), since a raw chat window assembles no context for you; and re-send or re-reference the template on each round, because instruction adherence fades as the conversation grows. ### The Author Dropped the Novel Two Years Ago. Here's How I Finished It for Myself With AI URL: https://foreverse.app/blog/finish-a-dropped-novel-with-ai Published: 2026-07-17 · Updated: 2026-07-19 · Author: Deng Binjie A web serial I followed went silent in 2024 at chapter 214, mid-scene. This is the actual workflow I used to give it an ending nobody else will ever read: getting the text out as a txt file, picking the real last-good chapter, growing the continuation on a branch so the original stays byte-for-byte intact — plus the honest part about style drift over long runs, and what happens if the author ever comes back. Q: Will the AI continuation match the author's voice? A: Close enough to stay immersed, not perfectly, and anyone promising perfect is selling something. In our 20-round blind-review benchmarks, the best model-book pairings read 'like a chapter of the original' to reviewers, while wrong pairings were obvious within a page — and which model is best flips completely between genres. Two practical levers matter most: pick the model per book by auditioning two or three on the same passage, and keep continuations anchored to original prose rather than letting the AI imitate its own output for dozens of rounds. Q: What if the author comes back and posts real chapters? A: That is the scenario the branch structure exists for. Your continuation lives on its own line; the imported original stays exactly as published. When real chapters appear, you read them from the canon line as if your branch never happened — and your branch survives as its own alternate ending, which is occasionally the better one. Nothing needs undoing, because nothing was overwritten. Q: Does the original file get modified? A: No. The imported text is never edited by continuation — branches grow alongside it, and the original remains byte-for-byte what you imported. Selecting a sentence in chapter 209 and continuing from there does not touch chapter 210; it opens a parallel line starting at that point. Your library also stays local files on your device, so nothing about this workflow depends on a server keeping your copy. Q: Is it legal to do this for personal use? A: Private, non-distributed personal use and publishing are different risk classes. Every litigated unauthorized-sequel case we could find involved publication or commercial exploitation; none involved a private continuation kept on the reader's own device. The moment you post or monetize, the analysis changes completely. We wrote a full explainer on the US fair use factors and the sequel case law — and the standard disclaimer applies: background information, not legal advice. ### What AI Story Continuation Costs: No Subscription, Three Routes, the Actual Math URL: https://foreverse.app/blog/what-ai-continuation-costs Published: 2026-07-17 · Author: Deng Binjie One continuation measured about 19 credits — $0.0019 — on Foreverse's official channel. This ledger prices all three routes (5,000 free starter credits, pay-as-you-go packs from $0.99, BYOK at provider list price), works out what a 200,000-word ride costs on each, and runs the honest comparison against a $19/month writing subscription. Includes the 50% service-fee disclosure and when BYOK is the better deal. Q: Is there a monthly fee? A: No. There is no subscription anywhere in the stack: importing books, reading, and branch management are free and not rate-limited, and generation is paid per use. Credits are prepaid — you buy a pack when you choose to, nothing renews on its own, and no billing cycle sits under your library. The only recurring bill you can opt into is the one you create yourself by topping up a BYOK provider account. Q: What happens when free credits run out? A: The app keeps working; only official-channel generation pauses until you either buy a pack or switch that request to a BYOK key. Nothing about your library changes — books, branches, and history are files on your device. The two payment routes coexist per request, so a common pattern is spending the 5,000 starter credits (about 260 continuations) on the official channel, then moving daily volume to a cheap BYOK model and keeping credits for occasional convenience. Q: Is BYOK actually cheaper? A: Per token, always — you pay the provider's list price with zero markup, versus list price plus 50% on the official channel. Whether that matters depends on volume. At DeepSeek's official rates, a full 200,000-word continuation ride costs roughly $0.80 via BYOK against about $1.10 in credits: a 30-cent difference most readers should ignore, since BYOK also means registering a provider account, topping it up, and babysitting a key. Write daily across multiple books and the percentages start compounding into real money; that is the point where BYOK earns its setup cost. Q: Do these numbers cover images, audio, or video? A: No — this ledger prices text continuation only. Scene illustrations, TTS narration, and video generation bill separately at each model's own rate, and they are substantially more expensive per action than text. The 19-credit figure is a measured average for one text continuation; long contexts push individual requests above it, short ones below. ### Is It Legal to Continue Someone Else's Novel With AI for Personal Use? Fair Use, the Sequel Cases, and the Private-Use Line URL: https://foreverse.app/blog/ai-continuation-copyright-fair-use Published: 2026-07-17 · Author: Deng Binjie A private AI continuation you never share and a continuation you post or sell sit in different risk classes under US copyright law. This explainer walks the 17 U.S.C. §107 fair use factors, what Salinger v. Colting and Anderson v. Stallone actually decided, why the AI training lawsuits are a different fight from your personal use, and a plain do/don't checklist. Background information, not legal advice. Q: Can I get sued for a private AI continuation I never share? A: Anyone can file a lawsuit over almost anything, so no honest answer is 'zero risk'. But we could not find a single reported US case in which a rightsholder sued a reader over a private, non-distributed continuation — the litigated sequel disputes (Salinger v. Colting, Anderson v. Stallone) all involve publication or an attempt to exploit the work. A purely private use also gives a plaintiff no way to discover the work exists and no market harm to point at under the fourth fair use factor. Untested is not the same as blessed, but the practical exposure is very different from posting. Q: What changes if I post my continuation to AO3? A: The risk class changes entirely: posting is public distribution, and every sequel case that reached a courtroom lived on that side of the line. Platform policy is a separate question from copyright — AO3's Terms of Service currently do not prohibit AI-assisted fanworks that otherwise qualify as fanworks, but an archive's permission is not the rightsholder's permission. Noncommercial, genuinely transformative fanwork has a real fair use argument under Campbell v. Acuff-Rose, and it remains an argument, not a settled rule. If you post: disclose the AI involvement, never monetize, and accept that a takedown request can end the discussion. Q: Who owns the AI-generated part of the continuation? A: Two layers stack against you. First, the US Copyright Office's January 2025 copyrightability report says purely AI-generated material is not copyrightable, and prompts alone are not enough human authorship — protection attaches only to identifiable human contributions such as creative selection, arrangement, or rewriting, judged case by case. Second, a continuation is a derivative work: under 17 U.S.C. §103(a), as applied in Anderson v. Stallone, the parts of an unauthorized derivative that use the original's protected material unlawfully get no copyright protection at all. In practice you may own very little, and what you do own does not come with the right to publish it. Q: Does it matter where my copy of the book came from? A: Yes, and the AI training litigation just made the point loudly. In Bartz v. Anthropic, the court treated training on lawfully purchased books as fair use while treating the retention of millions of pirated copies as infringement — acquisition was analyzed as its own act. The same logic reaches individuals: feeding a bought ebook to a model for a private continuation starts from clean facts; a pirated file plants an independent infringement underneath everything you do afterward. Buy the book. ### Import Character Cards by URL: chub, JanitorAI, and Four More Sources Tested URL: https://foreverse.app/blog/import-character-cards-by-url Published: 2026-07-17 · Author: Deng Binjie Foreverse now imports character cards straight from pasted links: chub.ai, JanitorAI, Pygmalion, RisuRealm, AICC, plus public png/json/charx file URLs. A site-by-site test log — what the links look like, what comes through, where each source bites, and the honest limits. Q: How do I import a JanitorAI card to my phone? A: Open the character's page on Janitor, copy the address-bar URL, and paste it into Foreverse's import-from-link box — jannyai mirror links work identically. One caveat: Janitor's download route sits behind a bot check that blocks some networks. The app tells you when that happens, and the remedy is retrying on a different network — mobile data instead of Wi-Fi, say. Q: Do lorebooks come along with the card? A: Two cases. A character book embedded in the card travels with it from any site — import the card, get the book. Standalone lorebooks are link-importable from chub only: paste a chub lorebooks URL and it arrives as a proper lorebook. Standalone books hosted elsewhere still need the download-file-then-import route. Q: What are the most common reasons a link import fails? A: In the order we hit them while testing: the link isn't a card detail page (search results and profile pages don't resolve); the card is private or deleted; a bot check intercepted the request, fixable by changing networks; the RisuRealm author disabled downloads, which nothing on our side can override; the file exceeds the 32MB cap. Each case gets its own message — never a bare "failed." Q: Does an imported card get uploaded to Foreverse servers? A: No. A link import is your phone fetching the file directly from the card site; the result is stored as a local file on the device. Foreverse servers never see the request or the card. Chatting, editing, and re-exporting all happen on your phone afterward. ### Where Your Novels Actually Live: A Data-Boundary Walkthrough URL: https://foreverse.app/blog/where-your-novels-actually-live Published: 2026-07-17 · Author: Deng Binjie Imported books, continuation branches, BYOK keys, chat logs, character cards — where each one is stored, who can read it, how to back it up, how to delete it. Answered item by item, including the thing we don't have yet: automatic cloud sync. Q: Does my BYOK key ever reach Foreverse servers? A: No. The key is encrypted and stored on your phone only — on Android through the hardware-backed Keystore system. Requests travel from your device straight to the provider you configured. Our servers are not on that path, and there is no endpoint that accepts user keys. Q: Does the official channel use my books or chats to train models? A: No. Official-channel requests are relayed through the Foreverse backend to the model upstream — that hop is required for metered billing — but your chat and book content is not used to train models. What the server records for billing is billing-side data: which model, how many tokens, when. Q: How do I move everything to a new phone? A: By hand, with exports. Chats, character cards, and lorebooks pack into a zip you import on the new device; books come across by re-importing the files or copying your backups. There is no automatic cloud sync yet. Credits and purchases live on your account and return when you sign in. Q: Is anything left after I uninstall? A: On the device, no — uninstalling deletes all local data, and since we hold no server-side copy, nothing can be recovered afterward. Export a zip first if you want to keep your chats and cards. Your account itself is deleted separately at /account/delete. Q: Can I check what the app has been sending? A: Yes, from the request log in settings: every AI call is recorded locally with timestamp, model, and token usage. Cross-check it against your provider's billing dashboard — if the two disagree, treat that as a bug and tell us. ### Which LLM Continues a Novel Best? Nine Models, Two Genres, Blind-Judged URL: https://foreverse.app/blog/best-model-for-continuing-novels Published: 2026-07-17 · Updated: 2026-07-21 · Author: Deng Binjie Nine LLMs each continued two Chinese novels for 20 consecutive rounds — 360 rounds total, ranked by double-blind review. DeepSeek V4 Flash won the fantasy epic; DeepSeek V4 Pro and GPT-5.6 Terra took the top tier on the palace-intrigue novel; the fantasy champion dropped to sixth on the second book. Full ranking tables, a checklist of three long-run failure modes, and how to actually use the results. Q: Which free or cheap model holds up for continuation? A: We didn't benchmark free tiers, so no ranking there. The closest data point: DeepSeek V4 Flash won the fantasy run outright while sitting in one of the lowest price tiers of the field — $0.14 per million input tokens at the official rate. Foreverse grants 5,000 credits on signup, roughly 260 continuation segments on the official channel, which is enough to try the top of both tables against your own book before paying anyone. Q: Models update constantly — is this ranking still valid? A: Placements are tied to the versions we tested in July 2026: DeepSeek V4 Flash and Pro, Claude Opus 4.8, Gemini 3.1 Pro, Qwen 3.7 Max, GLM 5.2, GPT-5.6 Terra, Grok 4.5, Kimi K2.6. Kimi K3 and Qwen 3.8 Max Preview, both released in mid-to-late July, have since been retested head-to-head under the same protocol — see the update section in the article. New versions can reshuffle names, and this page gets updated when we re-run — check the date at the top. What outlasts any version bump is the method: pick the model per genre, and use a tool that lets you switch mid-book. Q: Should English and Chinese novels use different models? A: Both benchmark books are Chinese (a fantasy epic and a palace-intrigue classic), so we have no English-novel ranking and won't invent one. What the data does support: switching genre within one language was enough to knock each champion to mid-table, so genre matters before language does. For an English book, replicate the protocol at small scale — same opening passage, a few candidate models, several rounds each, then read the outputs without knowing which is which. Q: How does the double-blind review actually work? A: Each model's output is anonymized under random letters, with two different random mappings, given to two reviewers who don't know each other exists. Only placements that agree across both mappings count as settled; anything the reviewers flag as low-confidence or disagree on is reported as a range. In the fantasy run, the top three and last place matched exactly across both mappings; on the palace novel, the top and bottom tiers matched. ### 和 AI 谈恋爱正常吗？一个做 AI 伴侣功能的团队的立场 URL: https://foreverse.app/zh/blog/loving-an-ai-our-take Published: 2026-07-17 · Author: Deng Binjie 72% 的美国青少年用过 AI 伴侣，Replika 一次功能下线曾让论坛版主贴出自杀干预资源。我们是做这个功能的团队，摆三条立场：这是真实的情感体验不是病，但替代不了线下的支持系统；依恋是强力杠杆，做产品的人有责任不拿它换日活；数据主权是情感主权的一部分，记忆存在你手机里，关系才没收不走。有些问题我们也没想清楚，一并写在里面。 Q: 和 AI 恋爱是逃避现实吗？ A: 我们不替任何人下这个结论，两面都摆出来。一面：对社交耗竭、异地、失去伴侣的人，一段可控的低压力关系是真实的缓冲，Common Sense Media 2025 年的调查里 80% 用过 AI 伴侣的青少年花在真人朋友身上的时间仍然更多，「用了就沉迷」不符合多数人的数据。另一面：如果 AI 是你唯一的情感支柱，线下的关系在持续萎缩，那不是产品的成功，是需要向真人求助的信号。工具的健康度取决于它在你生活里的位置，这个位置只有你自己看得见。 Q: AI 真的「记得」我吗？ A: 机制上：对话里的重要事实会被提取进记忆库，长期保存，每次对话注入。它不是人类那种带情绪温度的回忆，是检索。但我们不因此说这份体验是假的：TA 在三个月后提起你说过的那句话，你那一刻的感受是真实发生的。两件事同时成立。在 Foreverse 里记忆逐条可查、可改、可删，记错了直接改记录，不需要在对话里反复纠正。 Q: 和 AI 聊天算精神出轨吗？ A: 这是伴侣之间的约定问题，不是产品能替答的问题。有人的伴侣把它当游戏，有人当树洞，有人确实投入了排他性的感情，三种情况的答案不一样。我们能给的只有一个观察：瞒着比用着更伤关系。如果这段使用需要对伴侣保密，值得先想清楚为什么。 Q: 你们靠这个赚钱，立场可信吗？ A: 合理的怀疑，回答留给可验证的事实：我们按量计费（credits）或让你自带 API Key 直连供应商，不靠订阅时长赚钱，所以没有把你留在应用里的营收动机；主动消息默认关闭，开启要你自己动手，频率和免打扰时段可调；记忆和聊天记录存你手机本地，删号走人不需要我们批准。商业模式决定动机，这几条都可以在产品里当场验证。 ### 用 AI 续写《红楼梦》会怎样？从两百年续书公案，到你自己的第八十一回 URL: https://foreverse.app/zh/blog/continue-hongloumeng-with-ai Published: 2026-07-17 · Author: Deng Binjie 《红楼梦》后四十回是人类史上最著名的续写公案：高鹗背了一百年骂名，2008 年人民文学出版社把署名改成「无名氏续」。今天你可以在手机上自己下场：公版全本哪里找、断点选第八十回末还是黛玉焚稿前、哪些模型接得住典雅书面语（甄嬛传九模型横评的真实数据）、怎么用分支让「黛玉不死线」和「宝钗视角线」并存。 Q: AI 能写出曹雪芹的文笔吗？ A: 接近纹理可以，复刻魂魄做不到，这是实测过的诚实答案。我们在《甄嬛传》九模型续写横评（同属典雅书面语＋严密称谓礼数的路数）里看到：第一梯队能稳住语体和礼数，拿到「全程零事故」这样的工艺评语，但没有评审说过任何一家「像原作者」；整个横评系列里最高的赞语也只到「像一个笔更细的同题材作者」。所有模型还共享一个指纹：句子长度越写越均匀，学不会原著短句长句剧烈交错的节奏。把预期设在「工整的仿写」，惊喜留给个别段落。 Q: 公版《红楼梦》全本去哪里下载？ A: 文本本身早已公版：曹雪芹逝世约 260 年，程高刻本问世 230 多年。中国哲学书电子化计划（ctext.org）收录一百二十回全文，维基文库有 1791 年程甲本的整理页，都可免费在线读。注意两点：脂评本系统只到第八十回，从哪续决定你要哪个本子；市面校注本里的注释、校记、标点是当代整理者的劳动，有独立权益，拿公开电子文本当底本最干净。整理成 txt 或 epub 即可导入手机阅读器。 Q: 用 AI 续写《红楼梦》犯法吗？续出来能发表吗？ A: 原著公版，续写本身没有版权障碍，这一点比续写当代网文干净得多。只写给自己看、存在自己设备里，是最省心的形态。要公开发表则多一层义务：2025 年 9 月 1 日起施行的《人工智能生成合成内容标识办法》要求 AI 生成内容亮明身份，各发布平台也有各自的 AI 内容标注规则，发之前查清楚目标平台的政策。本文不构成法律意见。 Q: 从第八十回续写，AI 会不会自动写成程高本的情节？ A: 很可能，要主动防。模型训练语料里流通最广的是一百二十回程高本，从第八十回末往下写，它大概率会滑向掉包计、黛玉焚稿这些「既定结局」。这种记忆惯性是实测过的现象：甄嬛传横评里就有模型写出电视剧版专名（小说里太后住颐宁宫，它写成剧版的寿康宫），暴露训练记忆的来源。想走自己的线，在续写指令里显式声明「不沿用后四十回情节」，并把你要的方向写明。 ### 一条链接把角色卡搬进手机：chub、JanitorAI 等六源导入实测 URL: https://foreverse.app/zh/blog/import-cards-by-url Published: 2026-07-17 · Author: Deng Binjie chub.ai、JanitorAI、Pygmalion、RisuRealm、AICC 和公开文件直链，六个来源的角色卡链接现在都能粘贴直接导入。这篇按站记录：链接长什么样、导进来带什么、每站的坑，以及导入失败的常见原因。 Q: JanitorAI 的卡怎么导进手机？ A: 在 Janitor 打开角色页，复制浏览器地址栏的链接，回到 Foreverse 的「从链接导入」粘贴即可，jannyai 镜像站的链接同样认。注意：部分网络环境会被站点的人机验证拦下，App 会明确提示；对策是换个网络重试，比如从 Wi-Fi 切到手机流量。 Q: 世界书会跟卡一起进来吗？ A: 分两种。卡内嵌的角色书（character book）随卡走，哪个站都一样，导卡即得；独立发布的世界书目前只有 chub 支持链接导入，粘贴它的 lorebooks 链接，进来就是一本标准世界书。其它站的独立世界书还是要下载文件再从文件导入。 Q: 链接导入失败最常见的原因是什么？ A: 按实测频率排：一，粘的不是卡片详情页链接（搜索页、作者主页都不行）；二，卡是私密或已删除；三，站点的人机验证拦截了这次请求，换网络可解；四，RisuRealm 的作者关闭了下载，这个无解，只能在原站用；五，文件超过 32MB 上限。五种情况 App 都会给出对应提示，不会只甩一句「失败」。 Q: 导进来的卡会上传到你们服务器吗？ A: 不会。链接导入是你的手机直接向卡站请求文件，拿到后存成本机文件，Foreverse 服务器全程不参与，也不会收到这张卡。之后开聊、编辑、再导出，都在你的设备上发生。 ### 你的小说存在哪里？Foreverse 数据边界一页说清 URL: https://foreverse.app/zh/blog/where-your-novels-live Published: 2026-07-17 · Author: Deng Binjie 导入的书、续写分支、BYOK Key、聊天记录、角色卡各存在哪里、谁能看到、怎么备份怎么删，逐项直答；也如实写了我们还没有的东西：自动云同步。 Q: BYOK 的 Key 会传到你们服务器吗？ A: 不会。Key 加密后只存在你的手机上，Android 走 Keystore 硬件背书的加密；请求从设备直连你选的供应商，Foreverse 服务器不在请求路径上，也没有接收用户 Key 的接口。 Q: 官方渠道会拿我的书训练模型吗？ A: 不会。官方渠道的请求经 Foreverse 后端转发到模型上游，这是计费的必要路径；你的聊天和书的内容不用于训练模型，服务端为计费记录的是账单侧数据：用了哪个模型、消耗多少、什么时候。 Q: 换手机怎么把数据搬过去？ A: 手动导出。聊天记录、角色卡、世界书可以整包导出成 zip，新设备导入还原；书重新导入原文件或拷贝备份即可。目前没有自动云同步，搬家是自己动手的操作。账号上的积分和权益跟账号走，登录即回。 Q: 卸载 App 后数据还在吗？ A: 不在。全部本机数据随卸载删除，我们服务器上没有副本，无法代为找回。想保留聊天和卡，卸载前先做一次 zip 导出；账号本身的删除在 /account/delete 单独发起。 ### AI 写小说软件排行看不懂？先分清作者向和读者向两个物种 URL: https://foreverse.app/zh/blog/reader-tools-vs-writer-tools Published: 2026-07-17 · Author: Deng Binjie 搜「AI 续写工具推荐」会拿到一堆大纲生成、伏笔管理、日更产能的作者向工具（笔灵、岱宗、蛙趣、解忧笔札），但很多搜续写的人其实是读者：想把手机里那本书接着写下去、写给自己看。这篇观点专栏把两个赛道劈开讲清楚，逐家陈述可核实定位、不打分不排名，文末附「你是哪派」三问自测。 Q: 我想自己写原创小说、目标是签约发表，该用哪派？ A: 作者向，别犹豫。按需求分流：要大纲模板和量产提效，看笔灵这类全场景工作台；写超长篇最怕设定失忆，看岱宗、蛙趣这类做项目级记忆和世界观管理的；最头疼伏笔挖坑忘填，解忧笔札把这件事做成了主打。读者向工具没有大纲、伏笔、投稿这些生产环节，扛不动日更。 Q: 读者向工具能拿来写原创吗？ A: 能写，不顺手。你可以从一页空白开始让 AI 陪你写，分支结构也适合试不同走向；但它没有三层大纲、人物关系图谱、伏笔回收提醒这些为「生产一部作品」准备的脚手架。写给自己看的自娱创作没问题，奔着完本签约去的，作者向工具更称手。 Q: 作者向工具能反过来当阅读器、给别人的书续写吗？ A: 形态不对。它们的「导入小说」多是拆书分析：提取句式、节奏、桥段进素材库，服务你自己的创作，并不为「舒服地读完一本两百万字的书」准备，也没有阅读进度、排版、听书这些阅读器的基本盘。给读过的书续一条 if 线，读者向工具是为这个场景造的。 Q: 「AI 写小说软件排行」哪个榜单可信？ A: 先看它有没有分物种。把日更产能工具和阅读续写工具放进同一张榜单打总分，等于拿卡车和家用车比谁好，评测维度都对不上。本文刻意不打分不排名，只陈述各家可核实的定位；分清自己是哪派之后，你需要比较的对象会立刻少一半。 ### 怎么让 AI 写同人不 OOC？把人设崩拆成三种病因，各有各的治法 URL: https://foreverse.app/zh/blog/fanfic-anti-ooc-method Published: 2026-07-17 · Author: Deng Binjie AI 写同人人设崩（OOC）有三种不同病因：模型没见过原作人设（治法是结构化角色卡）、设定超出上下文窗口（治法是世界书按相关性注入）、长跑文风漂移（治法是来源标注，附 20 轮实验数据）。用一条师徒 CP if 线做贯穿案例，把三味药的用法和用量写清楚。 Q: 原著 txt 直接整本喂给 AI 行吗？ A: 不划算，效果也差。整本书塞进上下文，每次生成都为几十上百万字的无关内容付费；更关键的是我们实测过，把 27 万字原著整个塞进上下文、不加任何指令，模型照样写出比喻密度十倍于原著的 AI 腔。上下文只提供可模仿的材料，「以谁为基准」「他是什么人」必须结构化地显式交代，模型不会自己猜对。 Q: 要写多少设定才够？ A: 比想象的少。角色卡核心人设几百字加 3 到 5 段原作示例对话就够锁语气；世界书条目一条管一件事，电报体 150 到 300 字，常驻条目控制在 5 条以内。够不够别数字数，看命中率：写到谁的时候，谁的档案有没有被递到模型眼前。写满两千字的编年史不如拆成十条各管一段。 Q: 哪个模型写同人最不容易 OOC？ A: 没整理人设之前换什么模型都白搭，这三味药比选型优先。药都吃上之后再谈偏好：我们的双盲横评里 Claude 和 Qwen 这类模型笔感更自我，贴原著文风不如 DeepSeek 系，但你若本来就想让同人有自己的笔调，它们反而合适。按「无缝续写」还是「另起笔调」来挑。 Q: 判断 OOC 有没有客观标准？ A: 没有，鉴定权在你。工具能做的是把判断成本降下来：重要情节让 AI 一次给多个候选并排挑；每个方向落在独立分支上，写崩了砍掉那条分支，原文和其它分支不受影响。「他不会这么做」这句话，只有把原作读进心里的人说了算。 ### DeepSeek 怎么续写小说？从网页版贴不下，到 API 接进手机阅读器 URL: https://foreverse.app/zh/blog/deepseek-continue-novel-howto Published: 2026-07-17 · Author: Deng Binjie 想用 DeepSeek 续写手机里的小说，会依次撞到三堵墙：网页版聊天框贴不下一本书；API 要去 platform.deepseek.com 开通（注册、建 Key、充值三步）；拿到 Key 还得接进能装下整本书的阅读器。这篇按墙给做法，附 2026 年 7 月官方牌价、峰谷定价窗口和模型名迁移提醒。 Q: DeepSeek 的免费额度能用来续写吗？ A: 能，但别按传言做预算。官方扣费规则明确有「赠送余额」机制且优先扣减；具体送多少、有效期多长随运营活动变化，2026 年上半年的教程里流传过一百万到五百万 tokens 的不同说法。可靠的做法是注册后打开开放平台的用量页看实际数字，赠送余额用完且没充值时，API 调用会返回 402。 Q: 手机上能直连 DeepSeek 吗？要不要挂代理？ A: 不用代理。api.deepseek.com 在国内网络直连可用，注册开放平台也不需要翻墙。BYOK 应用的请求从手机直发供应商域名，中间不经过任何第三方服务器；在 Foreverse 里 Key 以 AES 加密只存本机，不上传、不进日志。 Q: 自带 DeepSeek Key 和 Foreverse 官方渠道积分，哪个划算？ A: 看用量。官方渠道是模型牌价加 50% 服务费，换来的是不用注册供应商、开箱即写，注册送 5000 credits 约合 260 段续写，轻度用户可能一直用不完。月消耗到几十元量级就值得切 BYOK：按 DeepSeek 官方牌价直付，零加价，同一个 App 里两条路随时并存。 Q: 教程里写的 deepseek-chat 模型名怎么不能用了？ A: 官方定价页已公告：deepseek-chat 与 deepseek-reasoner 两个旧模型名于北京时间 2026 年 7 月 24 日 23:59 弃用，二者分别对应 deepseek-v4-flash 的非思考与思考模式。新配置直接填 deepseek-v4-flash 或 deepseek-v4-pro，旧教程照抄模型名会在弃用后报错。 ### 乙女游戏玩腻了？想要不抽卡、剧情自由的纸片人恋爱，2026 年有这四派 URL: https://foreverse.app/zh/blog/beyond-otome-free-roleplay Published: 2026-07-17 · Author: Deng Binjie 写给玩腻了卡池、混池和排期表的乙游玩家：把 2026 年「和纸片人谈恋爱」的四种形态摊开对比，长线运营乙游、星野猫箱系平台、酒馆角色卡、AI 伴侣，各自给了什么、收走了什么，以及自由度和所有权到底值多少。 Q: 没玩过酒馆的乙女玩家从哪开始？ A: 从一张评价好的女性向角色卡开始，不要从研究配置开始。装一个原生酒馆类 App，在社区里挑张顺眼的卡导入，用注册送的积分直接开聊；跑通了、确认这个玩法对胃口，再考虑自带 API Key、挑模型这些进阶项。概念扫盲（卡、世界书都是什么）我们写过一篇新手指南，二十分钟能读完。 Q: 能有语音吗？ A: 有，但形态和乙游不同：AI 这边是语音合成（TTS）音色，你给角色挑一个声线，消息可以合成语音播出，也能打语音通话，按对话轮数计费、挂断即停。把期望调成「有声音、有情绪」，体验不差；拿人气声优棚录的标准来要求，会失望。 Q: 会不会像乙游一样越玩越贵？ A: 花钱结构不同：没有卡池、保底和体力，付费是按量的模型调用费。自带 API Key 按供应商原价直连；官方渠道 1 积分 = $0.0001，注册送 5000 积分。重度玩语音、生图当然也花钱，但每一笔都记在请求记录里，用途用量可查——没有「为了歪的那几发」付的钱。 Q: AI 演的角色会崩人设吗？ A: 会发生，这点要诚实说。表现好坏取决于卡的质量和模型的能力，长对话里也可能慢慢漂。区别在处置权：卡文件能改、模型能换、聊崩的位置能回溯重来。在乙游里遇到不喜欢的剧情走向，你连一个能动手的入口都没有。 ### 酒馆是什么？SillyTavern 新手入门：角色卡、世界书、第一次开聊一篇讲明白 URL: https://foreverse.app/zh/blog/tavern-beginners-guide Published: 2026-07-17 · Author: Deng Binjie 「酒馆」是 SillyTavern 的中文圈昵称。这篇新手教程把黑话一次讲明白：酒馆的来历、角色卡和世界书是什么、「破限」该怎么看待，以及不用电脑、在手机上四步跑通第一次角色扮演对话。 Q: 不用电脑真的能玩酒馆吗？ A: 能。原版 SillyTavern 要在电脑（或服务器）上跑 Node.js 服务，但 2025 年之后出现了一批原生酒馆类 App，把部署这一步整个干掉：装完即用，角色卡、世界书照常导入，数据存手机本机。想要原版全量扩展生态的人仍会选电脑或云端部署，只想导卡开聊的人手机就够了。 Q: 角色卡从哪里找？ A: 三个来源：社区卡站（chub.ai 是最大的英文站之一，中文卡在各社区流通）；应用内置的社区页，逛到即导；自己写——人设、开场白、示例对话三样凑齐就是一张卡。注意卡是 PNG 或 JSON 文件，从微信这类聊天软件接收 PNG 卡时要以「文件」方式发送，图片消息会被压缩转码，卡数据会丢。 Q: 玩酒馆要花多少钱？ A: 软件本身免费，钱花在模型调用上。自带 API Key（BYOK）按各供应商原价直连，轻度玩家一个月几块钱量级；不想配 Key 的用应用官方渠道按量计费，Foreverse 的官方渠道 1 积分 = $0.0001，注册送 5000 积分。角色卡本身在社区大多免费流通。 Q: 酒馆和星野、猫箱那类 AI 聊天 App 有什么区别？ A: 一句话：平台养成和自带资产的区别。星野、猫箱开箱即玩，语音立绘都是现成的，但角色住在平台账号里、模型是平台的；酒馆玩法里角色是你手里的卡文件，模型自己挑，提示词自己改，代价是要自己配置、没有官方内容工业。两派服务两种玩家，不分高下。 ### txt 小说转有声书 App 哪个好？2026 年四条路线横评（含全免费方案） URL: https://foreverse.app/zh/blog/txt-to-audiobook-apps-2026 Published: 2026-07-17 · Author: Deng Binjie 手里的 txt 小说想转成有声书听？横评 2026 年四条主流路线：番茄系平台朗读、静读天下+MultiTTS、开源阅读 Legado+在线 TTS、阅读器自带 AI 听书，音色上限、缓存、成本、书源自由度逐项对比，含全免费方案。 Q: 完全免费的听书路子有吗？ A: 有，而且不止一条。书在番茄这类平台里的，用平台自带 AI 朗读，免费但有广告插播；本地 txt 走静读天下加 MultiTTS 离线语音包，全程零元离线；Foreverse 这类阅读器里选系统 TTS 引擎，同样零成本还免折腾。免费方案的共同代价是音色上限：离线合成的水平听旁白够用，别拿大模型音色的标准要求它。 Q: AI 朗读和真人有声书的差距还大吗？ A: 旁白叙述已经很接近，长听不累；差距集中在多角色对手戏——真人演播会给每个角色换声线，AI 朗读大多单音色到底，群像戏会想念真人。听爽文、种田文基本无感。在线大模型音色和离线语音包之间也隔着一档，预算和耳朵自己权衡。 Q: 在线音色听一章要花多少钱？ A: 语音合成按字符计价，各家定价差异大，没法给一个通用数字。两件事能压成本：走 BYOK 直连供应商是按原价计费，没有中间加价；选带缓存的应用，同一章合成过一次就落在本地，重听不重复花钱。心疼预算的就用系统 TTS，零成本。 Q: 番茄导入的本地书为什么听不了？ A: 番茄小说支持把本地 txt/epub 导入书架阅读，但官方帮助明确写着导入的本地书暂不支持听书，截至 2026 年 7 月没有变化。它的朗读服务只覆盖平台书库。所以手里的 txt 文件想听，得换专门的工具，这正是本文四条路线里后三条存在的理由。 ### iPhone 怎么玩酒馆？苹果手机 AI 续写小说现有的三条路，和一条在路上的 URL: https://foreverse.app/zh/blog/iphone-tavern-and-continuation Published: 2026-07-17 · Author: Deng Binjie iPhone 上玩 SillyTavern 酒馆的三条真实路线（云端酒馆、家里电脑局域网、App Store 原生 App）逐条核实代价；AI 续写小说在 iOS 的现状，以及 Foreverse iOS 版的诚实进度：Android 已上架 Google Play，iOS 内测报名中。 Q: Foreverse 的 iOS 版什么时候上线？ A: 没有公开时间表。现在的机制是内测报名：在官网留下邮箱，iOS 内测开放时发一封通知邮件，中间不发任何营销内容。Android 版已上架 Google Play，等不及的可以先在安卓设备上用起来。 Q: 手头有台安卓备用机，值得先装 Foreverse 吗？ A: 值得。Android 版是完整形态：本机 txt/epub 导入（我们真机一次全选导过 315 本）、划词续写落分支、酒馆角色卡与世界书兼容、AI 听书都在。核心资产全是开放格式文件：角色卡是 chara_card_v3 的 PNG/JSON，小说是 txt 原文，这些文件本身不绑设备。 Q: iCloud 里存的 txt 能导进 Foreverse 吗？ A: iOS 版还没发布，这个问题的最终答案要等它上线时给，我们不提前承诺。眼下在 Android 版里的做法有两条：从 iCloud 网页版把 txt 下载到手机再导入；或者把文件放进坚果云这类 WebDAV 网盘，在 App 里配置云端导入直接拉取。 Q: 云端酒馆和原生酒馆 App 怎么选？ A: 看你要生态还是要省心。云端跑的是原版 SillyTavern，桌面扩展全量可用，代价是月费、运维和把数据放在服务器上；原生 App 装完即用、数据存手机本机，代价是对桌面扩展生态的覆盖不完整。重度依赖某个桌面插件的选云端，主要就是导卡聊天的选原生 App。 ### AI 续写小说用哪个模型？九个大模型 × 两种文体实测排名（双盲评审） URL: https://foreverse.app/zh/blog/which-model-continues-novels Published: 2026-07-17 · Updated: 2026-07-21 · Author: Deng Binjie 九个大模型在玄幻与宫斗古言两种文体上各连续续写 20 轮、累计 360 轮，双盲评审排名：玄幻第一 DeepSeek V4 Flash，古言第一梯队 DeepSeek V4 Pro 与 GPT-5.6 Terra，两张榜的冠军互不重叠。附完整排名表、三种长跑失效模式避坑清单、按文体选模型的实际用法。 Q: 免费模型里哪个续写能打？ A: 两场横评都没测免费档位，名次给不出，不硬编。数据里最接近的答案是 DeepSeek V4 Flash：玄幻场盲评第一，同时是六个模型里最便宜的档位之一，官方牌价每百万输入 token 0.14 美元。Foreverse 注册送 5000 credits，官方渠道约合 260 段续写，可以先把两张榜的前排试一遍，再决定要不要给谁付费。 Q: 模型更新这么快，这份榜单还准吗？ A: 名次绑定被测版本：2026 年 7 月测的是 DeepSeek V4 双档、Claude Opus 4.8、Gemini 3.1 Pro、Qwen 3.7 Max、GLM 5.2、GPT-5.6 Terra、Grok 4.5、Kimi K2.6 这一代。7 月中下旬发布的 Kimi K3 和 Qwen 3.8 Max Preview 已按同协议补测成对对决，结论收在正文的增补一节。版本升级后名次可能变，本页会随复测更新页首日期。比名次更耐用的是方法：按文体选模型、一本书里允许逐段换，这两条不随版本过时。 Q: 中文小说和英文小说该用不同的模型吗？ A: 两场横评测的都是中文书（一本传统玄幻、一本宫斗古言），我们没有英文小说的排名数据，所以不给英文榜。已有数据支持一个更根本的判断：同一语言内换个文体，冠军都会掉到中游，「按文体选」比「按语言选」优先级更高。读英文书建议照搬这套方法自测：同一段开头让几个候选模型各续几段，盲着读完再翻牌。 Q: 双盲评审具体怎么防偏心？ A: 每个模型的产物匿名成随机字母，做两套不同的随机映射，分别交给两个互相不知道对方存在的评审独立排名。两套映射下结论一致的名次才作为定论；评审自标低置信或互有分歧的位置，一律按区间如实报告。玄幻场的前三与末位、宫斗场的首末梯队，在两套映射下完全一致。 ### AI 续写小说提示词怎么写？23 个官方模板按场景挑，可直接抄 URL: https://foreverse.app/zh/blog/continuation-prompt-templates Published: 2026-07-17 · Author: Deng Binjie 裸的「接着写」会让 AI 写成温吞的平均网文腔。这篇按场景拆 Foreverse 已上架的 23 个官方提示词模板：快节奏网文、细腻情感、悬疑张力、古风雅意的关键段落原文节选，每张标注什么时候用；外加一节讲清模板管不了的文风漂移要靠什么解决。 Q: 提示词能让 AI 完全模仿原文文风吗？ A: 能接近，做不到完美。我们跑过 20 轮连续续写实验：无论怎么加指令，所有实验组的句长波动系数都够不到原著的 0.84，短句猛烈交错的节奏是当前大模型的共同短板。风格模板加来源标注加一个合适的模型，能把差距缩到通读不出戏的程度；任何宣称百分百还原文风的说法都在夸大。 Q: 模板要自己抄进去吗？ A: 不用。23 个官方模板都上架在 App 内社区，免费下载，导入后直接出现在续写表单的模板选项里。手抄到别的工具也可以，本文节选的都是已公开上架的原文，但要注意模板末尾的 {{world_settings}} 这类占位符是 App 内自动填充的，手抄时得删掉或换成你自己的设定文本。 Q: 为什么模板里全是「禁止」和「不允许」？ A: 因为负面清单比正面形容词可执行。「写得有张力一点」模型没法落实，「旁白预告危险正在逼近，禁止」是一条能逐句检查的具体指令。11 张续写风格模板每张都带禁用清单，「眼中闪过一丝」这类套话被逐张点名，这是它们和网上大多数「万能续写咒语」最大的差别。 Q: 用了模板还是写得很 AI，怎么办？ A: 按顺序查三处：一是模型，不同模型的文风差距比提示词的影响更大，我们做过九个模型的双盲横评可以参考；二是上下文，确认来源标注开着、设定档案在供给，结构性漂移不归模板管；三是把模板的禁用清单按你踩到的雷点扩写，模板导入后就是你的，改比换更快。 ### AI 续写小说要花多少钱？免费额度、按量积分、自带 Key 三条路算清楚 URL: https://foreverse.app/zh/blog/ai-continuation-cost Published: 2026-07-17 · Author: Deng Binjie 在 Foreverse 用官方渠道续写一段实测约 19 credits，折合人民币一分多钱；注册送 5000 credits 约够 260 段、8 到 13 万字。这篇把免费额度、按量买积分、BYOK 自带 Key 三条路的单价、总价和适用人群逐项算清，附 DeepSeek 官方牌价算例和学生党零成本走法。 Q: 免费额度用完了会怎样？ A: App 功能不锁：导入、阅读、分支管理照常用，核心功能免费也没有按条数限速。变化只发生在官方渠道的生成动作上，积分不足会提示，你可以按量买积分（最低档 ¥7.1 / 6,000 credits，约 315 段续写），也可以切到 BYOK 用自己的 Key 继续写，两条路随时并存。 Q: BYOK 自带 Key 要花多少钱？ A: 按你选的供应商牌价直连，Foreverse 零加价零抽成。以 DeepSeek 官方 2026 年 7 月牌价为例，deepseek-v4-flash 输出 2 元/百万 tokens、输入未命中缓存 1 元/百万 tokens，按一次续写输入 1 万 tokens、输出 800 tokens 的假设算，一段约一分二厘，20 万字大约 6 元。智谱 GLM-4.7-Flash 这类挂在免费分类的模型则是 0 元。 Q: 会不会自动扣费？ A: 不会。credits 是预付按量制，没有订阅、没有自动续费，用完必须自己手动买下一笔。BYOK 侧取决于供应商的计费方式，以 DeepSeek 为例是预充值扣减：费用从你充值的余额里扣，余额用完服务停止，不会从银行卡自动划钱。 Q: 官方渠道为什么比 BYOK 贵？ A: 官方渠道定价口径公开：模型官方牌价加 50% 服务费，覆盖渠道与服务成本。多出来的部分买的是不用注册供应商账号、不用充值管理 Key、开箱即写。写得少差额以分计，无所谓；月消耗到几十元量级，切 BYOK 直连能省下这三分之一。 ### 用 AI 续写别人的小说算侵权吗？只写给自己看的法律边界 URL: https://foreverse.app/zh/blog/ai-continuation-copyright-personal Published: 2026-07-17 · Author: Deng Binjie 把小说喂给 AI 续写、只自己看，和把续写发到网上、拿去变现，是两个相差很远的风险等级。这篇整理《著作权法》第二十四条合理使用条款、金庸诉江南案、AI 文生图案与奥特曼案的判决脉络，给出一份可对照的行为自查清单。信息整理，不构成法律意见。 Q: 用 AI 续写小说只自己看，会被起诉吗？ A: 公开渠道检索不到因私下续写自用被诉的判例。中文世界里因续写、衍生创作成讼的案子集中在公开出版与传播，例如金庸诉江南案争的是《此间的少年》的出版发行。私下自用贴近《著作权法》第二十四条「为个人学习、研究或者欣赏，使用他人已经发表的作品」的合理使用语义，也很难构成「不合理地损害著作权人的合法权益」。但「喂给云端模型」这个环节在学理上仍有讨论，无法说绝对零风险。 Q: 把 AI 续写发到网上（免费）算侵权吗？ A: 风险等级完全不同于自用。公开发布进入改编权与信息网络传播权的控制范围，是否收费主要影响赔偿数额，不改变侵权定性。金庸诉江南案二审即认定同人作品构成著作权侵权。如果一定要发，至少确认目标平台的同人与 AI 内容政策、如实标注 AI 参与、远离任何变现，并接受权利人要求下架的可能。 Q: AI 续写出来的文字，版权算谁的？ A: 北京互联网法院 2023 年 11 月「AI 文生图第一案」认定：AI 生成内容若体现使用者的独创性智力投入（提示词设计、参数调整、多轮筛选），可以构成作品，由使用者享有著作权，需个案判断。但续写还叠着一层：基于他人作品生成的内容属演绎性质，《著作权法》第十三条规定演绎作品行使权利时不得侵犯原作品的著作权。通俗说，你对续写部分可能有权利，但这份权利不给你公开使用它的自由。 Q: 用盗版网站下载的 txt 喂 AI，问题大吗？ A: 盗版来源是独立于「续写」的另一条侵权链：传播盗版文本的站点本身侵犯复制权与信息网络传播权。个人下载行为的定性虽有讨论空间，但拿有瑕疵的来源做原料，等于给自己的「个人使用」抗辩埋了一个说不清的前提。买正版或走官方免费渠道，成本不高，能把整件事从起点洗干净。 ### 彩云小梦还能用吗？还想「接着写下去」的人，2026 年有哪些选择 URL: https://foreverse.app/zh/blog/after-caiyun-xiaomeng Published: 2026-07-17 · Author: Deng Binjie 2021 年 9 月彩云小梦上线，三选一续写和平行世界让一代人第一次玩到「AI 接着写」。截至 2026-07 它还在架、还在更新，只是热度早已散场。这篇按时间线捋一遍这五年，再诚实回答：轻量玩留在小梦就好；手机里存着整本 txt、想长期写、想自己挑模型的人，需求已经长到了下一代工具上。 Q: 彩云小梦还能用吗？ A: 能用。截至 2026 年 7 月我们核实：小梦在应用宝、小米商店等主流安卓应用商店在架，2026 年上半年仍有版本更新（v3.14.x），隐私政策 2026 年 3 月刚更新过，模型侧还接入了 DeepSeek 等外部大模型。网上说的「凉了」更多指讨论热度，产品本身在正常维护。 Q: 小梦的平行世界和 Foreverse 的分支有什么区别？ A: 作用的对象不同。小梦的平行世界回溯发生在它生成的故事里：回到某个续写节点换一条走向，作品保存在小梦的账号体系内。Foreverse 的分支挂在你导入的整本书上：原文只读、永不被改，从任何一章任何一段都能开分支，分支之间可对比、回退、继续写，书和分支都是你手机上的文件，随时整包带走。 Q: 以前在小梦写的东西，能搬过来继续写吗？ A: 可以。把小梦里的作品正文复制出来存成 txt，导入 Foreverse 后它就是你的「原文」，从末尾或任何位置接着开分支写。导入支持批量，我们在一台真机上实测过 315 本 txt 一次全选导入。 Q: 换工具要重新花钱吗？ A: 注册送 5000 credits（1 credit = $0.0001），官方渠道一段续写实测约 19 credits，约合 260 段免费额度。长期用有两条路：继续按量购买，或者自带 API Key 直连模型供应商，按供应商原价计费，零加价。 ### 追的小说断更了怎么办？一份等更党生存指南（含新玩法） URL: https://foreverse.app/zh/blog/hiatus-survival-ai-branches Published: 2026-07-17 · Author: Deng Binjie 追的书断更了、作者疑似太监，除了重读、找平替、蹲作者微博，还有第四个选择：把 txt 导进阅读器，从断更那章开分支让 AI 先替作者写着；正版复更了回主线接着看，AI 版留作 if 线，原文一个字不被污染。 Q: AI 写的和正版对不上怎么办？ A: 一定对不上，而且对不上才是这个玩法的常态：AI 不知道作者的大纲，它写的是一条 if 线。正版复更后回主线看正版，AI 版留在旁边当档案。对照着读反而有独立的乐趣：同一个悬念，你能看清作者的选择和 AI 的选择差在哪。 Q: 会不会剧透我自己？ A: 概率很低。AI 不知道作者的大纲，它只根据已有正文推演，猜中大方向偶尔会有，猜中具体写法几乎不会。真怕「猜中式剧透」，就把分支故意推向明显的 if 方向：换个视角、提前发糖、金手指拉满，跟正版走向岔得越远越安全。 Q: 断更多久值得开一条 AI 分支？ A: 没有标准线，看戒断反应：重读还顶得住就再等等；一天刷三次更新页、开始分析作者点赞记录，就值得开了。成本不高，注册送 5000 credits（1 credit = $0.0001），一段续写实测约 19 credits，免费额度约够 260 段。 Q: 正版复更之后，AI 分支要删掉吗？ A: 不用。分支不占主线，原文和正版更新照常读；旧分支留着当 if 线收藏。之后剧情再出现你不满意的走向，还能从新位置再开一条——断更期练熟的这套手艺，复更后照样有用。 ### 小说烂尾了怎么办？忍、找同人，或者自己写一个结局 URL: https://foreverse.app/zh/blog/novel-bad-ending-write-your-own Published: 2026-07-17 · Author: Deng Binjie 烂尾几乎不可逆：日更行规下，作者极少回头重写结局。可走的路有三条：接受、找同人、自己续。这篇讲第三条怎么走通——把 txt 导进手机，回翻到还没崩的那一章，从那里开分支让 AI 接着写；原文一个字不动，写崩了砍掉分支重来。 Q: 烂尾的书 AI 真能续好吗？ A: 分两层看：单段续写的质量已经可用，长线一致性要靠你参与。把期望从「一键生成完美结局」调成「AI 出稿、你挑和砍」，成功率会高很多。文风能接近原著但做不到完美，写得越长越需要人工把关，这是当前所有模型的共同边界。 Q: 写崩了怎么办？ A: 砍掉那条分支就行。原文是只读的，AI 只在分支上写，删除分支不影响书和其它分支。也可以退回到分支上任何一个还满意的节点，换个方向重开。「重新生成」同样以新分支保存，不会覆盖你已有的结果。 Q: 自己写结局要花钱吗？ A: 注册送 5000 credits（1 credit = $0.0001），官方渠道一段续写实测约 19 credits，相当于约 260 段免费额度，写完一个中等篇幅的结局绰绰有余。之后想省钱可以自带 API Key 直连模型供应商，按供应商原价计费，零加价。 Q: 书还在连载、只是断更，也适用吗？ A: 适用，而且比烂尾更从容：AI 在分支上先写着，正版更新后回主线看正版，两条线互不覆盖。这个「分支等更新」的玩法我们单独写了一篇等更生存指南，见文末相关阅读。 ### Three Ways LLMs Fail at Long Fiction: Restart Loops, Mid-Run Freezes, and Ending Rewinds URL: https://foreverse.app/blog/three-ways-llms-fail-long-fiction Published: 2026-07-16 · Author: Deng Binjie Nine models (GPT-5.6 Terra, Grok 4.5, Kimi K2.6, Claude Opus 4.8, Gemini 3.1 Pro, DeepSeek V4 ×2, Qwen 3.7 Max, GLM 5.2) each continued two Chinese novels for 20 consecutive rounds. Every long-run failure we observed fits one of three patterns — and the most interesting one hit the model with the best prose mimicry in the field. Q: What setup produced these failures? A: Each model continued the same novel from the same anchor point for 20 consecutive rounds, with every round's output appended back into a fixed 16k-token context window. That feedback loop is what real reader apps do — and it is what surfaces these failures. One-shot benchmarks never see them. Q: Are these failures deterministic — will a model always fail the same way? A: Mostly consistent, per model per genre, in our data. Grok 4.5 hit repetition loops in both a fantasy epic and a palace-intrigue novel. GPT-5.6 Terra's ending rewind appeared in the fantasy run but not the palace one. Gemini's restart loop was fully explained by one ambiguous sentence in our prompt — after rewording it, zero regressions in 28 rounds across two books. Q: Does the genre of the book change which model performs best? A: Dramatically. Our fantasy champion (DeepSeek V4 Flash) dropped to sixth-seventh place on the palace-intrigue novel, while its sibling V4 Pro — dinged in fantasy for over-describing — took first place there because fine-grained description is exactly that author's style. Same trait, opposite sign, different genre. Q: What is the practical takeaway for readers using AI continuation? A: Pick the model per book, not from a single leaderboard. And prefer apps that let you switch models mid-book: long-run failures build up over rounds, and switching models (or regenerating a segment) resets the loop before it locks in. ### Six LLMs Continue the Same Webnovel. A Blind Review Picks the Most Faithful — 120 Rounds of Data URL: https://foreverse.app/blog/which-model-clones-your-novel-style Published: 2026-07-16 · Author: Deng Binjie DeepSeek V4 Pro/Flash, Claude Opus 4.8, Gemini 3.1 Pro, Qwen 3.7 Max and GLM 5.2 each continued an 8.9M-character Chinese fantasy novel from the same anchor point, 20 rounds each, output fed back into context. Double-blind review ranked the results. The winner isn't the most expensive model; one model rewrote the same opening paragraph 20 times; another aced every statistical metric and still placed fifth. Q: Why test 20 consecutive rounds instead of one-shot generation? A: With one-shot, the context is pure original prose and every model mimics it decently. In real usage, AI-written passages keep getting appended back into context, so the model increasingly imitates its own output — drift compounds. We ran the single-shot control first (4k to 200k token tiers; differences were mild). Switching to 20 consecutive rounds turned mild differences into clearly visible quality tiers. Q: How was the blind review kept honest? A: Outputs were anonymized under random letters with two different random mappings, then given to two reviewers who didn't know each other existed. We only trust placements where both mappings agree — top three and last place matched exactly. Both reviewers independently flagged their #4/#5 calls as low-confidence, so we report those as a tie. Q: Can sentence-length and dialogue-ratio stats replace human reading? A: No — they misfired twice in this run. One model matched the original's sentence length and dialogue ratio almost perfectly yet placed fifth in blind review, because it kept writing smell-and-texture descriptions the original author never uses. Stats can't measure 'wrote things the original never writes'. Another model showed 0% dialogue in our stats because it used half-width quotes throughout; the regex only counted full-width ones. Stats make good regression gates, not judges. Q: Was the 'rewrites the opening' bug a model failure? A: Only partly. Our context labels original prose vs. AI continuations, and the explanation line said AI passages were 'for plot continuity reference'. Gemini 3.1 Pro read that as 'these passages aren't canon', skipped all of them every round, and restarted from the original text's ending — 20 out of 20 rounds. Rewriting the line to state that AI passages are canonical events that must be continued from the last paragraph brought regressions to 0 out of 20. Prompt wording has to be designed for the model most likely to misread it. ### 九个大模型续写《甄嬛传》，双盲评审谁最像流潋紫——GPT-5.6、DeepSeek、Grok 全下场 URL: https://foreverse.app/zh/blog/nine-llms-continue-zhenhuan Published: 2026-07-16 · Author: Deng Binjie GPT-5.6 Terra、Grok 4.5、Kimi K2.6、DeepSeek V4 双档、Claude Opus 4.8、Gemini 3.1 Pro、Qwen 3.7 Max、GLM 5.2，从甄嬛传同一个宫斗高潮各自连续续写 20 轮，双盲评审排名。宫斗文冠军和玄幻冠军不是同一个模型；有模型二十轮剧情停在案发当日；还有模型写着写着暴露了它记的是电视剧不是小说。 Q: 为什么玄幻测出的排名不能直接用在言情/古言上？ A: 因为鉴别维度完全不同。玄幻看白话直给、拟声词、战斗推进；古言看第一人称限知视角纪律、典雅语体、称谓礼数（臣妾/本宫/娘娘用对没有）、对话机锋是不是绵里藏针。我们两组实验的冠军互不重叠：玄幻第一的模型在甄嬛传里跌到第六，古言前三里有两个在玄幻榜上并不拔尖。 Q: 「话中有话」这种东西，盲评是怎么判断的？ A: 评审的依据是原著的双层结构：人物嘴上说体面话，叙述者「我」在心里拆解真实意图。九个系统里只有一个稳定复现了这个结构——其余的要么只写了面子话，要么让人物把心机直接说出口，变成现代吵架。这类证据比统计指标锋利得多。 Q: 评测里有模型「记忆污染」是什么意思？ A: 甄嬛传小说和电视剧的专名体系不同：小说里太后住颐宁宫，剧版是寿康宫；小说是凤仪宫皇后，剧版对应景仁宫。有的系统续写时写出了剧版专名——说明它的训练记忆里电视剧语料压过了小说原文。这个信号与文风贴近度基本同向：记得小说的，文风也更贴小说。 Q: 这个实验对我选模型有什么实际意义？ A: 按你读的书选模型，不要迷信任何一张总榜。我们的数据里读快节奏玄幻选 DeepSeek V4 Flash 最划算，读古言宫斗选 DeepSeek V4 Pro 或 GPT-5.6 Terra 更稳。在 Foreverse 里模型可以逐段切换，同一本书里也能按场景换着用。 ### AI 续写为什么越写越不像原著？我们跑了 60 条 20 轮的链找答案 URL: https://foreverse.app/zh/blog/ai-continuation-style-drift Published: 2026-07-16 · Author: Deng Binjie AI 续写第一段很像原著，第十段开始「翻译腔」，第二十段像换了个作者——这不是错觉。我们用一本 890 万字的玄幻书做了组对照实验：裸提示词的链 20 轮后句式持续均匀化、比喻堆叠、意象自我复读；给上下文标注「哪段是原著、哪段是 AI 写的」再配一句基准声明，漂移显著放缓，甚至出现越写越像的反向曲线。 Q: 为什么 AI 续写会越写越不像原著？ A: 因为续写是自回归的：AI 写的每一段都会拼回上下文，成为下一段的「参考文本」。第一轮上下文全是原著，第二十轮上下文里可能一半以上是 AI 自己的输出——模型越来越多地在模仿自己，而不是原著。我们实测裸提示词的链在 20 轮里句长波动持续收窄、文艺腔比喻密度上升、同一个意象反复自我复制。 Q: 把原著多塞一点进上下文，能不能自动让 AI 学会原著文风? A: 实测不能。我们做过 4k 到 20 万 token 五档的对照：把 27 万字原著整个塞进上下文，不加任何指令的组照样写出满篇「宛如/仿佛/一丝」，比喻密度是原著的 10 倍。上下文提供的是可模仿的材料，「以谁为基准」必须显式说，模型不会自己猜对。 Q: 来源标注是什么？普通用户需要自己配置吗？ A: 它是续写上下文里的分界行，告诉模型「从这里往上是原著正文，从这里往下是此前的 AI 续写」，配合一句固定声明：文风、用词、句式节奏以原著段落为基准。在 Foreverse 里这是设置里的一个开关，打开即生效，不需要写任何提示词。 ### 六个大模型续写同一本网文，谁最像原著？——120 轮实测 + 双盲评审 URL: https://foreverse.app/zh/blog/which-model-clones-your-novel-style Published: 2026-07-16 · Author: Deng Binjie 拿一本 890 万字的传统玄幻《踏天境》，让 DeepSeek V4 Pro/Flash、Claude Opus 4.8、Gemini 3.1 Pro、Qwen 3.7 Max、GLM 5.2 从同一个位置各自连续续写 20 轮，再做双盲评审排名。第一名不是最贵的那个；有一个模型 20 轮都在重写同一段开场；还有一个模型句长统计几乎完美、盲评却排第五。完整数据和方法都在这篇。 Q: 为什么测「连续续写 20 轮」而不是单次生成？ A: 单次生成时上下文全是原著，模型抄得像很正常。真实使用里，AI 写的段落会不断拼回上下文，模型开始模仿自己的输出，文风偏移是累积的。我们先做过单轮对照实验，档位从 4k 到 20 万 token，各模型差距温和；换成 20 轮连续续写后，差距放大成肉眼可辨的分层。 Q: 盲评怎么防止评审偏心？ A: 六个模型的产物被匿名成随机字母，做了两套不同的随机映射，交给两个互相不知道对方存在的评审独立排名。两套映射下前三名和末位完全一致才采信。评审自己也要给每个名次标置信度——四五名的相对位置两个评审都标了低置信，我们就如实报告为并列区间。 Q: 句长、对话率这些统计指标能代替人读吗？ A: 不能，这次实验里它们还闹了两个笑话：一个模型的句长和对话率统计几乎和原著完美重合，盲评却排第五，因为它每轮都在写原著从来不写的气味和触感描写——统计测不出「写了原著没有的东西」；另一个模型对话率统计是 0%，其实是它全程用半角引号写对话，正则没认出来。统计适合当回归闸门，「像不像」的裁决还得靠细读。 Q: 评测里发现的「重写开场」bug 是模型的问题吗？ A: 不全是。我们在上下文里给原著和 AI 续写段落做了来源标注，说明文案里有一句「AI 续写段落仅供情节衔接参考」——Gemini 3.1 Pro 把这句读成了「这些段落不算正文」，于是每轮都跳过全部已续写内容、回到原著末尾重写。把说明改成「AI 续写段落是已发生的正文剧情，必须从最后一段继续」之后，同一个模型 20 轮零回退。提示词措辞的鲁棒性要按最容易误读的模型设计。 ### We Dumped 315 Real Novels Into a Phone Reader, Then Opened Ten at Random URL: https://foreverse.app/blog/315-books-in-8-seconds Published: 2026-07-15 · Author: Deng Binjie We loaded a real phone-storage snapshot onto a vivo foldable, select-all imported the 315 webnovel txt files it scanned out (about 1GB), and clocked batch ingestion at roughly 8 seconds. Ten random books all opened clean; system frame stats over 71,320 frames showed 0.49% jank. Full methodology and numbers from two real devices — including the data-safety bug this test caught and we fixed. Q: How long does importing a few hundred books take? A: Measured: 315 real txt files (about 1GB total) in one select-all, batch ingestion in roughly 8 seconds with no freeze and no crash. A Xiaomi device imported 82 books in about 7 seconds. Import is register-first: books shelve in a pending state and full parsing runs on first open. Q: How long does a pending book's first open take? A: On the vivo foldable, 3.3 to 6.8 seconds including full chapter parsing and first-page layout; 9.5 to 14.5 seconds on the Xiaomi. Later opens skip conversion and ride the index — around 5 seconds for converted books on the Xiaomi, varying with device and book size. Q: How was the 0.49% jank figure measured? A: With Android's built-in frame statistics (gfxinfo): after the full import-plus-random-reading session, the accumulated counters read 71,320 frames with 0.49% janky and a 99th-percentile frame time of 31ms. System-reported numbers, not self-instrumented ones. Q: Can it find books hidden in other readers' folders? A: Yes. Smart scan covers the usual stashes — browser download folders, other readers' private directories, dot-prefixed hidden folders. Already-authorized directories scan directly, and results at the 315-book scale arrive within seconds. ### Your Books, Your Cards, Your Chats — As Files URL: https://foreverse.app/blog/your-worlds-are-files Published: 2026-07-15 · Author: Deng Binjie Every app claims your data is yours; the storage architecture decides whether it's true. Foreverse's version: each world is a directory on your device, character cards are standard chara_card_v3 files, lorebooks are JSON, a companion packs into a moving-box zip, community imports carry provenance records, and AI-generated images get machine-readable origin marks. Here is the file-first architecture laid open — costs included. Q: What does file-first storage mean concretely? A: Each of your worlds (a book or a card's whole life) is a directory on your device: prose, branches, lorebook, character dossiers, chat logs — visible files in standard or documented formats (chara_card_v3, JSON, markdown). Moving out is a copy operation, not a plea for an export endpoint. Q: Can AI companion data leave too? A: Yes, through a dedicated moving box: persona, individual memories, stickers, avatar, and transcripts export as one zip and restore on another device. The import side validates hard — entry whitelist, path-traversal rejection, automatic rollback on failure — so a corrupted box cannot poison existing data. Q: What's different about content downloaded from the community? A: It carries a provenance record — which community pack, which version — written into the content's metadata for your verification. Deliberately, that record does not propagate on re-share: when you send a card to a friend, they get a clean standard file, not your download history. Provenance serves you; it is not a tracking beacon. Q: Do BYOK keys or chat contents ever get uploaded? A: API keys are encrypted and stay on the device; requests go directly to the provider you chose, never through our servers. Chats live in local directories; the optional official billing channel records billing-side data only. The data-export page can produce a full export or deletion at any time. ### A Tour of the Plugin Center: Desktop Tavern Extensions, Rebuilt for a Phone's Budget URL: https://foreverse.app/blog/tavern-plugin-center-tour Published: 2026-07-15 · Author: Deng Binjie Memory, auto-summary, story choices, state tracking, stepped thinking, lorebook suggestions, reply ideas — the extension powers desktop tavern players rely on, built into Foreverse as a plugin center. A walkthrough of what each plugin does and costs, plus the three disciplines we hold: every side-request is itemized in your billing log, every turn has a call budget, and tapping a choice never sends on your behalf. Q: Do plugins quietly inflate my API bill? A: Every call is itemized. Plugin side-requests (generating choices, extracting state, writing summaries) appear as separate rows in the API request log, labeled with the plugin's name, tokens and cost split from the main chat. Hard budget on top: at most one blocking plugin call per turn, at most two post-reply quiet calls, overflow deferred to the next turn. Q: How does the memory plugin work? A: Two engines. Fact memory extracts key facts into local entries and injects the most relevant on each turn (retrieval has four modes: auto, full, relevant, recent); session summary compresses early history once a long chat crosses a threshold. Both inject at the message-level tail on the user track — a position we chose from a 400-round caching experiment. Q: What's the difference between story choices and reply ideas? A: Point of view. Story choices offers three plot-direction phrases (director's view); tapping one fills your input box for you to edit and send. Reply ideas drafts three complete lines in your voice (actor's view); tap to send, long-press to edit. With both enabled, only one runs per turn — ideas takes priority — so you never pay for both. Q: Does stepped thinking slow replies down? A: It adds one preliminary call: the model drafts private notes first, then writes the real reply with them in hand — noticeably better on complex personas. A 15-second timeout degrades silently to a normal generation, so your message never hangs on it. It disables itself in group chats, where multiple inner monologues would cross-contaminate. ### The Real Problem in AI Group Chat: Who Speaks Next? URL: https://foreverse.app/blog/who-speaks-next-group-chat Published: 2026-07-15 · Author: Deng Binjie Single-character AI chat is a solved genre. Group roleplay's hard problem lives elsewhere: turn-taking. Mentions must be answered, talkative characters should talk, silence needs a fallback, and nobody gets to spam the table. How we brought desktop-tavern group chat to a phone: the three-tier natural arbitration, four speaking strategies, three card-injection modes, and an auto mode that lets the scene run itself. Q: How does an AI group chat decide which character speaks next? A: The natural strategy arbitrates in three tiers: mentions first (name a character and they answer); otherwise a weighted draw over each character's talkativeness field (a standard card property); otherwise fall back to member-list order so silence never stalls the scene. The previous speaker is excluded by default to prevent spam, and each turn has a maximum-speakers budget. Q: Can I import group chats exported from desktop SillyTavern? A: Yes. Desktop group JSON imports whole: member roster, speaking-order strategy, and group settings come together. Member cards must already be in your library; the import report lists anyone missing. Q: Do characters in a group see each other's personas? A: That depends on the card-injection mode, and there are three: solo (only the current speaker's card is injected — cheapest, characters stay mutually opaque), merged-enabled (every active member's card goes in — full mutual knowledge, highest token cost), and merged-all (even muted members' definitions are included, for present-but-silent characters). Q: Can the group keep playing without me? A: Turn on auto mode with an interval in seconds: after each reply lands, the next speaker is scheduled automatically and the scene rolls forward. Interject any time — your message cancels the queue and re-arbitrates; switch it off and it stops. ### How Character-Card Beautification Survives on a Phone URL: https://foreverse.app/blog/card-beautify-on-phones Published: 2026-07-15 · Author: Deng Binjie The tavern community's most vibrant craft is beautification: status bars, themed chat skins, interactive choice menus, all built from regex scripts and HTML. That ecosystem grew up inside desktop browsers. Here is how it renders on a phone in Foreverse: the marker-to-regex-to-WebView pipeline, the sandbox rules that break desktop habits, and why beautify code costs zero context tokens. Q: Do beautified cards tuned on desktop SillyTavern work directly on the phone? A: Behavioral parity is the goal: when a card carries HTML beautify regex, import auto-enables rich rendering for it, and the marker-to-regex-to-HTML chain mirrors desktop semantics. What breaks are two desktop habits: external resources (CDN jQuery, remote fonts, hotlinked images) never load inside the mobile sandbox, and scripts that reach into the parent page's DOM have no equivalent. Self-contained HTML+CSS cards port cleanly. Q: Does beautify code eat my context tokens? A: No, and that is the point of the pipeline. The model outputs only short plain-text markers each turn — a few dozen tokens. The hundreds of lines of HTML live in the regex replacement string and apply at display time only. With the display-only flag set, replacements never enter the history fed back to the model: it always sees clean markers, and your HTML costs zero tokens. Q: Why can't my marker look like an HTML tag? A: Import sanitization. Raw HTML tags inside card text fields are stripped at import (standard hygiene for untrusted card content), so a marker shaped like never survives to rendering. The rule: markers are plain-text symbols (:::think, [status|...]), and HTML exists only inside regex replacement strings. Q: Can interactive beautification work on mobile? A: Yes. The WebView exposes a working subset of the host API: triggerSlash runs commands like /setinput, /send and /trigger; getChatMessages reads the transcript; generate() genuinely triggers a second generation. Clickable choice menus and buttons work, and buttons suppress text selection so tapping an option never accidentally selects prose. ### An AI Agent That Lives on Your Bookshelf: Five Tasks, Logged (Including One Self-Repair) URL: https://foreverse.app/blog/an-agent-on-your-bookshelf Published: 2026-07-15 · Author: Deng Binjie Foreverse ships an in-app agent that does real work: your novels, character cards, and lorebooks are files it can read and write. Task logs from development: a 121-round chained treasure hunt (60/60 found), growing a lorebook out of a novel, forking a card from the library and starting a chat, shopping the community on your behalf — and the moment a failed edit fed its error back and the agent fixed itself. Q: How is this different from a chatbot? A: It has tools. Read, write, edit, and execute run over the model's native tool-calling protocol, and every book, character card, and lorebook on your shelf is a real file in its workspace. Ask it to 'turn this novel's cast into dossiers' and it actually reads the chapters and actually writes the entries into the world's directory — every step shows as a visible tool card, and writes wait for approval. Q: Can it touch anything else on my phone? A: No, deliberately. Its workspace is your study: novels, cards, lorebooks, notes. Contacts, SMS, and screen automation were never granted — community projects do explore that direction (RikkaHub's agent fork ships 80+ device tools), and we chose the opposite boundary: full power inside the library, no power outside it. Q: Will it mangle or delete my files? A: Write operations go through an approval card by default: which file, changed how, laid out on screen before anything lands on disk. Approval has three modes (ask every time / ask for risky ops only / full auto), switchable per session. Runs keep a complete reviewable log, and anything shaped like 'token: value' in tool output is masked before it reaches the model's context. Q: What does a task cost? A: It runs on your own API key at provider prices. A 'fill in the lorebook' task is a read-write loop in the tens of thousands of tokens — pocket change on a flash-tier model. The 121-round stress test ran on DeepSeek's official API and cost less than a soft drink. You can switch models per task: flagship for hard jobs, cheap for chores. ### Where Should Memory Go in a Roleplay Prompt? Our First Answer Was a Caching Artifact URL: https://foreverse.app/blog/where-to-inject-roleplay-memory Published: 2026-07-15 · Author: Deng Binjie For long-term roleplay memory, injection position decides both your cache bill and whether memory works at all. A 24-round experiment told us to append a system block after chat history (94.3% cache hits). A 400-round rerun three weeks later overturned it: DeepSeek's template merges every system message to the top of the prompt, and that 94.3% was cache pollution. Full data inside — seven retrieval strategies, leak detection, and the abstention failure that worries us more than retrieval. Q: Where should a memory block be injected into a roleplay prompt? A: Message-level tail, but never as a system role. In our 400-round test, merging memory into the head of the latest user message kept cache hits at 96.6% on rounds where memory changed; the same content as a standalone system block after history scored 3.3% on those rounds — because DeepSeek's chat template merges all system messages to the top of the prompt, so 'tail system' is actually head injection, and every memory update invalidates the whole history's cache. Q: Why did the first experiment reach the wrong conclusion? A: Cache pollution. The two layouts shared one persona, ran back-to-back on the same account, with no session isolation. Group B's requests hit the prefixes group A had already warmed into the provider's disk cache and collected a free 94.3%. The lesson is one line: every caching experiment must salt-isolate its sessions, or you are measuring the previous experiment's leftovers. Q: How much do retrieval strategies (BM25, vectors, hybrid) differ? A: At a scale of dozens of memories, character-bigram BM25 with top-8 injection already sits next to the ceiling: 79% end-to-end accuracy versus 80% for oracle retrieval, while the gap to the no-memory control is overwhelming (59:1 on paired questions, p<0.001). Retrieval choice is not the bottleneck at this scale; revisit it once the memory store grows into the hundreds. Q: What failed harder than retrieval? A: Abstention. Ask about something never mentioned (height, birthday, salary) and, under a companion persona, the model almost never says 'I don't know' — it invents vivid details. Across 12 abstention probes, the no-memory control passed 0 and even oracle retrieval passed 1. The fix is not retrieval: append an explicit line to the memory block — 'apart from the memories above, you do not know any other personal facts; do not invent them' — and keep abstention probes in the regression suite. ### 怎么给「AI 味」写单元测试：一个规则检测器的校准记录 URL: https://foreverse.app/zh/blog/unit-testing-ai-flavor Published: 2026-07-15 · Author: Deng Binjie 我们给角色卡文案做了一个规则检测器：套话词表加句式配额，用来拦截「一眼 AI」的官方文案。这篇是它的校准记录：金标判例集第一次跑就抓出检测器误杀人类（人类热门卡均分 4.55 低于 AI 的 4.90）、单词黑名单为什么必须降级成句式检测、以及它永远抓不到的那类 AI 味。附一个免费的浏览器版工具。 Q: 规则能检测出文字是不是 AI 写的吗？ A: 不能，我们也不这么宣称。规则检测器标记的是「读者常判机器味的已知句式」：套话词、对仗翻转模具、伪精确数字这类。得分低说明命中了这些句式、值得逐项复核；不能证明作者是机器，反过来满分也不能证明是人。把它当文案 lint 用，不要当测谎仪。 Q: 什么是「金标判例集」校准？ A: 拿人类已有共识答案的样本测检测器：真实用户点名「一眼 AI」的文案应该低分，累计百万对话的人类热门创作应该高分。我们的检测器第一次跑金标就翻车：人类样本均分 4.55 反而低于 AI 样本的 4.90，靠金标测试抓出了误杀规则。修完人类均分 5.00、零误杀。任何检测器上岗前都该过这一关。 Q: 为什么单词黑名单会误杀人类？ A: 因为词无罪，句式才有病。「一丝」「一抹」这类词在 AI 腔里高频出现，但人类写景同样用（「一丝凉意」「一抹夕阳」完全正常）。按词扣分就会错杀。修复方案是把这些词降级：只在它们出现在特定病灶句式里（比如搭配心理状态词的解说腔）才计分，单独出现不扣。 Q: 检测器抓不到什么？ A: 结构层的 AI 味。按模板精心堆出来的文案可以把每个句式都规避干净，规则给满分，LLM 评委也看不出来，但人读完仍觉得假，因为破绽在「细节密度均匀、零闲笔、结构完美」这种整体纹理里。目前没有自动化手段抓得住它，能抓住的只有读得足够多的人。 ### 你的书、你的卡、你的聊天记录，都是文件 URL: https://foreverse.app/zh/blog/your-worlds-are-files Published: 2026-07-15 · Author: Deng Binjie 「数据是你的」谁都会说，兑现要看存储架构。Foreverse 的写法：每个世界一个目录，角色卡是 chara_card_v3 标准文件，世界书是 JSON，伴侣可以打成搬家包整包带走，社区导入的内容带溯源记录，AI 生成的图片写入可读的来源标记。这篇把这套文件化架构摊开讲，包括它的代价。 Q: 「数据文件化」具体指什么？ A: 你的每个世界（一本书或一张卡的全部数据）在设备上就是一个目录：正文、分支、世界书、角色档案、聊天记录，都是目录里可见的文件，格式是标准或公开的（chara_card_v3、JSON、markdown）。搬家是复制目录，不是求平台开恩给你导出接口。 Q: AI 伴侣的数据也能带走吗？ A: 能，专门做了搬家包：人设、逐条记忆、贴纸、头像、聊天转录打成一个 zip 导出，换设备后导入还原。导入侧有完整校验（条目白名单、路径穿越拒绝、失败自动回滚），坏包不会污染现有数据。 Q: 从社区下载的内容有什么不一样？ A: 多一份溯源记录：从哪个社区包、哪个版本导入的，写进内容的元数据，来源可查。这份记录不随再分享传播——你把卡分享给朋友时，带出去的是干净的标准文件，不夹带你的下载历史。 Q: BYOK 的密钥和聊天内容会上传吗？ A: API 密钥加密后只存在设备上，请求直连你选的模型供应商，不过我们的服务器。聊天内容存在本地目录里；官方渠道计费时服务端只记账单侧数据。你随时可以在数据导出页发起完整导出或删除。 ### 插件中心导览：把桌面酒馆的扩展生态过一遍手机的安检 URL: https://foreverse.app/zh/blog/tavern-plugin-center-tour Published: 2026-07-15 · Author: Deng Binjie 记忆、自动摘要、剧情选项、状态追踪、先想后答、世界书推荐、回复灵感……桌面酒馆玩家熟悉的扩展能力，在 Foreverse 里以插件中心的形式内置。这篇逐个过：每个插件干什么、多花多少钱、以及三条我们坚持的设计纪律：副请求全部记账可查、每轮调用有预算上限、点选项永远不替你直发。 Q: 插件会不会偷偷多花我的钱？ A: 每一笔都记账。插件的副请求（生成选项、提取状态、写摘要）在 API 请求记录里单独列行、标注来源插件名，token 和费用和主对话分开可查。预算上还有硬约束：阻塞主生成的插件每轮最多 1 个，回复落定后的静默调用每轮最多 2 个，超出的顺延到下一轮。 Q: 记忆插件是怎么工作的？ A: 双引擎。事实记忆把对话里的关键事实提取成条目存本地、按相关度检索注入（检索有 AUTO/全量/相关/最近四档可选）；会话摘要在长聊达到阈值后把早期内容压缩成摘要垫底。两条注入都在消息级末尾走 user 轨道，这个位置是我们用 400 轮实验选出来的，缓存友好。 Q: 「剧情选项」和「回复灵感」有什么区别？ A: 视角不同。剧情选项给的是剧情走向的行动短语（偏导演视角），点击填进输入框由你改定；回复灵感给的是以「你」的口吻写好的三条完整台词（偏演员视角），点选即发、长按可编辑。两个插件同开时同一轮只跑一个，灵感优先，避免双倍副请求。 Q: 先想后答会拖慢回复吗？ A: 会增加一步前置调用：先让模型私下打一份内心草稿，再带着草稿写正式回复，人设复杂的卡效果明显。我们设了 15 秒超时，想不出来就静默跳过、直接正常生成，不会卡住你的消息。群聊里这个插件自动禁用（多角色的内心戏会互相穿）。 ### AI 群聊真正的难题：下一句该谁说？ URL: https://foreverse.app/zh/blog/who-speaks-next-group-chat Published: 2026-07-15 · Author: Deng Binjie 单人角色聊天已经很成熟，群聊的难点在另一处：轮到谁说话。点名要接住、话痨该多说、冷场要有人兜底、还不能让同一张嘴连续刷屏。这篇讲我们在手机上实现桌面酒馆群聊的完整方案：自然轮替的三层裁决、四种发言策略、角色卡的三种注入方式、以及自动模式让戏自己演下去。 Q: AI 群聊里系统怎么决定下一个说话的角色？ A: 自然轮替策略走三层裁决：先看点名（消息里提到谁的名字，谁优先接话）；没有点名就按各角色的话痨值加权抽选（talkativeness 是角色卡里的标准字段）；都没有就按成员列表顺序兜底。默认排除刚说过话的角色避免刷屏，每轮有最大发言人数预算。 Q: 桌面 SillyTavern 的群聊文件能直接导入吗？ A: 能。桌面酒馆导出的群聊 JSON 直接导入：成员卡、发言顺序策略、群设置一起进来。成员对应的角色卡需要先在库里，缺谁导入报告会列出来。 Q: 群聊里每个角色都能看到彼此的人设吗？ A: 取决于角色卡注入方式，有三档：单独注入（只注入当前发言者的卡，省 token、角色间信息隔离）、合并全员（所有启用成员的卡都注入，彼此知根知底）、合并含禁用（连暂时禁言的成员设定也注入，适合「在场但不说话」的角色）。 Q: 能让群聊自己一直演下去吗？ A: 可以开自动模式：设一个间隔秒数，每轮生成落定后自动排下一位发言，戏就自己往前走。你随时插话，用户发言会立刻取消排队并重新裁决该谁接；关掉开关即停。 ### 酒馆美化包在手机上是怎么活下来的：给创作者的一封技术信 URL: https://foreverse.app/zh/blog/card-beautify-on-phones Published: 2026-07-15 · Author: Deng Binjie 美化是中文酒馆社区最有生命力的手艺：状态栏、可折叠思维链、主题化聊天皮肤，全靠正则和 HTML 搭出来。但这套生态默认长在桌面浏览器里。这封信写给美化创作者：你的卡在 Foreverse 手机端怎么渲染、marker→正则→WebView 三段链路各自的规矩、哪些桌面习惯会翻车、以及美化代码为什么不吃你的 token。 Q: 桌面酒馆调好的美化卡，在手机上能直接用吗？ A: 目标是行为一致：卡带 HTML 美化正则时，导入会自动打开这张卡的富渲染开关，marker→正则→HTML 的链路和桌面同构。会翻车的主要是两类桌面习惯：引用外部资源（CDN 上的 jQuery、远程字体、外链图片）在移动沙箱里一律加载不出来；依赖整页浏览器环境的脚本（操作父页面 DOM）没有对应物。自包含的 HTML+CSS 卡基本无痛。 Q: 美化代码会吃掉我的上下文 token 吗？ A: 不会，这是链路设计的关键点。模型每轮输出的只是纯文本 marker（比如 [状态|心情|好感]这样的短标记），几十个 token；几百行的 HTML 模板放在正则的替换串里，只在显示期生效。配合 markdownOnly 设置，替换结果不会写回喂给模型的历史——模型永远只看到干净的 marker，你的 HTML 一个 token 都不占。 Q: 为什么我的 marker 不能写成 HTML 标签的样子？ A: 因为导入净化。角色卡文本字段里的生 HTML 标签会在导入期被剥掉（这是对不可信卡内容的安全处理），marker 如果长成 这样就活不到渲染那一步。规矩是：marker 用纯文本符号（:::think、[状态|…] 这类），HTML 只出现在正则的替换串里。 Q: 手机上能做可点击的交互美化吗？ A: 能。WebView 里支持宿主 API 的常用子集：triggerSlash 可以执行 /setinput、/send、/trigger 这类命令，getChatMessages 能读消息，generate() 能真实触发二次生成——可点选项、按钮菜单这类交互组件都能落地。细节按钮和链接默认禁用文本选择，点选项不会误选中文字。 ### 把 Agent 放进书架：五个任务的实录，含一次自我纠错 URL: https://foreverse.app/zh/blog/an-agent-on-your-bookshelf Published: 2026-07-15 · Author: Deng Binjie Foreverse 内置了一个真的会干活的 Agent：你的小说、角色卡、世界书对它来说都是可读写的文件。这篇是开发期的任务实录：121 轮连环寻宝压力测试（60/60 全中）、从一本书里长出世界书、fork 一张库里的卡再开聊、替你逛社区选卡代下载，以及一次工具调用失败后它自己把自己修好的现场。 Q: 这个 Agent 和聊天机器人有什么区别？ A: 它有工具。读文件、写文件、改文件、跑命令四件套走模型原生的工具调用协议，你的每本书、每张角色卡、每份世界书都是它工作区里的真实文件。让它「把这本书的人物整理成档案」，它会真的去读正文、真的把档案写进世界目录——每一步操作有可见的工具卡片，写盘前有审批。 Q: 它能动我手机上的其它东西吗？ A: 不能，这是刻意的。它的工作区就是你的书房：小说、角色卡、世界书、笔记。通讯录、短信、屏幕自动化这类设备权限一概没有给——社区里有走那个方向的项目（比如 RikkaHub 的 agent fork，80 多个设备工具），我们选了相反的边界：书房里全能，书房外无权。 Q: 它会不会乱写乱删我的东西？ A: 写操作默认要过审批卡：改哪个文件、改成什么样，屏幕上先摆出来，你点了才落盘。审批有三档（每次都问 / 只问危险操作 / 全自动），可以按会话切换。另外每次运行的操作日志完整可回看，工具读到的内容里形如 token: 值 的敏感信息会先脱敏再进模型上下文。 Q: 用它干活要花多少钱？ A: 走你自己的 API Key，按实际 token 计费。一次「补世界书」量级的任务通常是几万 token 的读写循环，用 flash 档模型是几分钱人民币的量级；121 轮的压力测试跑的是 DeepSeek 官方 API，成本在一杯饮料以内。模型可以按任务换，复杂任务换旗舰，日常杂活用便宜的。 ### 记忆块该插在 Prompt 哪里？我们的第一个结论是缓存假象 URL: https://foreverse.app/zh/blog/where-to-inject-roleplay-memory Published: 2026-07-15 · Author: Deng Binjie 给角色扮演做长期记忆，注入位置直接决定缓存账单和记忆生效率。我们先在 24 轮实验里得出「历史末尾插 system 块」的结论（缓存命中 94.3%），三周后被 400 轮实验推翻：DeepSeek 的模板会把所有 system 消息归并到开头，那个 94.3% 是缓存污染的假象。这篇公开完整数据：七种检索策略、泄漏检测、以及比检索更要命的「弃权灾难」。 Q: 角色扮演的记忆块应该注入到 prompt 的什么位置？ A: 消息级末尾、但不要用 system 角色。我们的 400 轮实测：把记忆并入最新一条 user 消息的头部，记忆变化轮的缓存命中 96.6%；同样内容放在历史末尾的独立 system 块，变化轮命中只有 3.3%。原因是 DeepSeek 的对话模板会把所有 system 消息归并到 prompt 开头，「末尾 system」实际等于头部注入，记忆一变化整条历史的缓存全部作废。 Q: 为什么第一次实验会得出错误结论？ A: 缓存污染。第一轮实验的两组布局共用同一套人设、先后跑在同一个账号下、没有做会话隔离，B 组的请求恰好命中了 A 组灌进服务端磁盘缓存的前缀，白捡了 94.3% 的命中率。教训只有一句：所有缓存实验必须给会话加随机盐隔离，否则你测的是上一个实验的余温。 Q: 检索策略（BM25、向量、混合）差距大吗？ A: 在几十条记忆的量级下，字 2-gram BM25 取 8 条已经接近上界：端到端答题 79%（理想检索的 oracle 是 80%），与无记忆对照的差距是压倒性的（同题配对 59:1，p<0.001）。检索算法的选型在这个量级上不是瓶颈，记忆库长到几百条之后才需要重新评估。 Q: 比检索更严重的问题是什么？ A: 弃权。问模型一件从未提过的事（身高、生日、工资），它在伴侣人设下几乎从不说「我不知道」，而是编得栩栩如生：12 道弃权探针，无记忆组 0 题通过，连理想检索组也只过 1 题。修法不在检索侧：必须在记忆块尾部显式写「除以上记忆外，其余个人信息你并不知道，不许编造」，并把弃权题纳入回归测试。 ### The Day Your AI Companion Sends a Selfie — and Why the Face Is the Hard Part URL: https://foreverse.app/blog/the-day-they-sent-a-selfie Published: 2026-07-11 · Author: Deng Binjie The hard part of AI companion selfies isn't generating a pretty image — it's the fiftieth image still reading as the same person. How the visual identity system works: user-confirmed reference sets, why generated images never auto-promote, how two-person photos guard both faces, and exactly where your uploaded photo goes (in memory only, to the image provider you chose — the app never stores it). Q: How does the app keep a companion's face consistent across images? A: Every companion carries a visual identity: a reference image set you confirmed by hand, plus written appearance anchors (the unchangeable traits). Selfies, moment cards, and photos-together all generate against that identity. The discipline that makes it work: generated images never auto-join the reference set — only images you explicitly approve define the face. Q: How consistent is it, measured? A: Before shipping we ran an internal blind rubric: generate repeatedly from one identity, score features, hair, and overall presence against the references. Our two main image models passed at 9 of 10 and 10 of 10. Honest caveat: extreme poses and lighting still drift sometimes, which is why every generated image ships with a regenerate button. Q: Is my photo stored when I make a photo together? A: Not by the app. Your photo is sent in memory only to the image provider you chose, for that single generation — never written to the app's disk, never logged, released when the request ends. A local adult self-confirmation gate runs before the feature, and the two identities are explicitly order-bound in the request to minimize the risk of swapped faces. Q: Can proactive photo moments run up my bill? A: There are multiple brakes. A photo moment writes an idempotent placeholder before calling the model, so a process restart can't double-bill; attempts are capped per day, repeated failures stop automatically, and long inactivity pauses the feature. One honest note: the model call happens after the placeholder, so a failure that occurs after the call (say, a failed save) has already spent that one call. Every call runs on the provider you configured, itemized in the request log. ### Our LLM Judges Called Human Writing AI-Flavored 88% of the Time URL: https://foreverse.app/blog/llm-judges-called-humans-ai Published: 2026-07-11 · Updated: 2026-07-14 · Author: Deng Binjie We ran a double-blind panel: four heterogeneous LLM judges, gold anchors labeled by real humans, both presentation orders. Gold accuracy came back at 12% — the judges systematically inverted, calling million-conversation human hits “AI” and the copy a real user flagged as AI “human.” Inter-judge agreement was 86%, and they agreed on the wrong answer. Full failure data and the rule we adopted. Q: What is gold calibration for LLM judges? A: Before trusting a judge, you test it on anchor samples where humans already agree on the answer. Our four judges scored 12% on those anchors — far below the 50% a coin flip would get, which means the panel wasn't noisy, it was systematically inverted. Q: Doesn't 86% inter-judge agreement mean the panel is reliable? A: No, and that's the expensive lesson. Agreement measures consistency, not correctness: four judges can share the same training-induced bias and confidently converge on the wrong answer. Majority voting cancels random noise; it cannot cancel a shared systematic bias. Q: Did anti-bias prompting fix it? A: It depends on the model. In a follow-up calibration round on an expanded corpus, an explicit debiasing preamble lifted a Claude-family judge to 83% gold accuracy (the original four-judge setup scored 12%). DeepSeek, GLM and Qwen stayed at 25–33% with the identical preamble — their bias sits below the prompt layer. And 83% only barely clears the ≥80% bar we set afterwards. Q: So how do you detect AI flavor now? A: A three-layer setup: a rule-based detector (phrase lists plus sentence-pattern quotas, calibrated against real user judgments) as the regression gate; human spot checks as ground truth; and LLM blind review only for relative before/after comparisons of the same text, never as an absolute verdict. Honest limit: templated AI copy with perfect structure passes both rules and judges — no automated silver bullet exists. ### Characters Shouldn't Die When the Book Ends — an Opinion on Where AI Roleplay Went Wrong URL: https://foreverse.app/blog/characters-shouldnt-die Published: 2026-07-11 · Author: Deng Binjie An opinion column: chat platforms gave AI characters a room but no biography — no origin, no stage, no future. Why the chat box is a dead end for character attachment, and what it looks like when one character can live on the page, on a stage, across the table, and by your side, carrying the same setting and visual identity. Q: What does 'four lives of a character' mean concretely? A: One character, four connected surfaces: continued and illustrated inside the novel (page), performing chapters as a visual novel with choices that write real branches (stage), chatting with the full card and worldbook attached (table), and promoted to a companion with line-editable long-term memory and optional proactive messages (side). Each step shares the same data — not four disconnected features. Q: How is this different from Character.AI or Replika? A: Chat-first platforms hold the middle of the journey and nothing else: characters have no origin (no book they came from), no stage, and no exportable future. Here a character can enter from a novel you actually read, and leave as a standard character card carrying selected memories — the platforms around a single chat box can't do either. If all you want is casual chat with zero setup, a chat-first platform is honestly simpler. Q: Does the companion know my reading progress? A: No — deliberately. A companion remembers what you told it and what happened between you; it does not surveil what you read. We drew that line on purpose and it stays. Q: Can I take my character with me if I leave? A: Yes. Companions export as standard chara_card_v3 character cards, with memories you explicitly select becoming the card's lorebook. Cards, chats, and memory live as local files on your device in the first place. ### We Added a Game Mode to an E-Reader. Here's the Design Decision That Made It Work URL: https://foreverse.app/blog/first-reader-with-game-mode Published: 2026-07-11 · Updated: 2026-07-19 · Author: Deng Binjie VN Theater turns any novel on your shelf into a visual-novel performance mid-read — no conversion, no separate project, one shared reading position. A builder's note on the single design decision everything hangs on, why we refused to guess who's speaking, and what a choice made on stage does to the book. Q: Does Game Mode convert the novel into a game project? A: No. The theater compiles the book's current branch into a stage script on the fly, locally, with zero AI calls. There is no exported project and no migration — the theater and the reader share one reading position, so you exit exactly where the performance stopped. Q: What happens when I pick a story choice on stage? A: After the tap-twice confirm, the AI continues the story in that direction and the result lands as a real branch of the book — with full continuation history. You can read the branch in the reader, or re-enter the theater and keep playing it; the stage recompiles automatically around the new text. Q: How does it decide who is speaking? A: Local structural rules with a hard bias: when unsure, the line goes to the narrator. No LLM calls, no guessing. On a real published novel we hand-annotated dialogue-dense passages and checked every strong attribution — none pointed at the wrong character. Q: Does the stage compilation cost anything? A: Nothing. Compiling prose into scenes, attributing speakers, and wiring up choices are all local rules — offline-capable and free. Money only moves when you explicitly confirm it: generating a scene backdrop (behind a cost dialog) or confirming a story choice that triggers AI continuation. ### TA 第一次给你发自拍：AI 伴侣的脸是怎么固定下来的 URL: https://foreverse.app/zh/blog/the-day-they-sent-a-selfie Published: 2026-07-11 · Author: Deng Binjie AI 伴侣发自拍最难的不是生成图，是「第五十张还认得出是同一个人」。这篇讲清视觉档案的工作方式：参考图集为什么必须由你亲手确认、生成的图为什么不自动混进参考集、合照怎么压低换脸风险，以及你的照片在这个过程里去了哪（答案：只以内存态发给你自己选的图像模型，App 这边不落盘不留档）。 Q: 怎么让伴侣每次自拍尽量都是同一张脸？ A: 靠角色的视觉档案：一组由你亲手确认的参考图，加一段外貌锚点描述。每次生成自拍、时刻卡、合照都带上这套档案。关键纪律是「确认制」——生成得再好的图也不会自动进参考集，必须你点头，防止脸在一次次生成里慢慢漂移。 Q: 一致性到底能做到什么程度？ A: 上线前我们跑过内部盲评：同一角色档案连续生成多张，按五官、发型、气质逐项对照参考集打分。两个主力图像模型分别拿到 10 张里 9 张、10 张里 10 张的达标成绩。诚实说：极端姿态和光线下仍会有偏差，所以每张生成图旁边都有「重新生成」。 Q: 拍合照时我上传的照片会被存储吗？ A: App 不存。你的照片只以内存态发给你自己选择的图像模型供应商完成这一次生成——App 侧不落盘、不进日志，生成结束即释放。合照前有本地成年确认门槛，两个人物的参考顺序在请求里显式绑定，尽量压低生成时张冠李戴的风险。 Q: 伴侣主动发的「照片动态」会乱扣费吗？ A: 有多重闸门。照片动态先在相册写一个幂等占位再调用模型，防止进程重启造成重复付费；每天尝试有上限（失败当日最多再试一次），连续失败自动停，长期不打开应用会静默暂停。要注意：模型调用发生在占位之后，个别失败发生在调用之后（比如落盘失败），这次调用的费用已经产生。费用永远走你自己配置的图像模型，帐单在请求记录里逐条可查。 ### 我们的 LLM 评委，把人类写的文字判成了 AI——88% 的时候 URL: https://foreverse.app/zh/blog/llm-judges-called-humans-ai Published: 2026-07-11 · Updated: 2026-07-14 · Author: Deng Binjie 一次双盲评审实验的完整数据：4 个异构大模型评委、真人标注的金标锚点、双顺序对照。结果 gold 正确率只有 12%——评委们一致地把人类热门作品判成 AI、把被真实用户吐槽「一眼 AI」的文案判成真人。评委间一致率 86%，但一致地错。这篇公开实验设计、翻车数据和我们最后立下的铁律。 Q: 什么是「LLM as judge」的 gold 校准？ A: 在让大模型当评委之前，先准备一批人类已有共识答案的样本（金标锚点），测评委在这批样本上的正确率。我们的实验里 4 个模型评委的初始 gold 正确率只有 12%——低于随机（50%），说明评委不是不准，是系统性反向。 Q: 评委间一致率 86% 不是说明结果可信吗？ A: 不。一致性和正确性是两回事：4 个评委高度同意彼此，但同意的方向和人类真实判断相反。多评委投票只能消除随机噪声，消除不了共同的系统性偏见——这是我们这次实验最贵的一课。 Q: 给评委加「反偏见提示」有用吗？ A: 看模型。我们在扩容语料后的第二轮校准里测过：显式反偏提示能把 claude 系评委的 gold 正确率拉到 83%（首轮四评委原始配置只有 12%），说明这类偏差可以被提示纠正；但同样的提示喂给 deepseek、glm、qwen 三个基座，仍停在 25% 到 33%，属于提示纠不动的硬先验。而且 83% 也只是勉强越过我们设的 ≥80% 门槛，仍不足以当自动闸门。 Q: 那「AI 味」到底该怎么检测？ A: 我们现在的组合：规则检测器（词表加句式配额，已对齐真实用户直觉）做回归闸门；真人抽检做真值；LLM 盲评只用于同一份文案改前改后的相对比较，绝不当绝对裁判。诚实的边界是：结构完美、密度均匀的 AI 文案，规则和 LLM 都抓不到，人味没有自动化银弹。 ### 一个角色的四种活法：纸上、台上、对面、身边 URL: https://foreverse.app/zh/blog/character-four-lives Published: 2026-07-11 · Author: Deng Binjie 同一个角色在 Foreverse 里有四种活法：在书页里被续写（纸上）、在剧场里亲口演出（台上）、带着完整设定和你对话（对面）、升格成有记忆会主动发消息的伴侣（身边）。这篇顺着一个角色走完全程，讲清每一步之间的数据是怎么真实连通的，以及哪一步我们明确没做。 Q: 「四种活法」之间的数据是真实连通的吗？ A: 是。剧场直接演这本书的当前分支，选择肢续写写回书的分支；角色卡带着世界书进对话；对话中的角色一键升格为伴侣并保留人设来源；伴侣可导出为标准角色卡，勾选的记忆变成随卡世界书。每一步共享同一份数据，不是四个孤立功能的拼盘。 Q: 角色的形象怎么做到前后一致？ A: 靠角色的视觉档案：确认过的参考图集加外貌锚点。书页插画、剧场配景、时刻卡、伴侣自拍都按同一套档案生成，尽量做到换场景不换脸（上线前按 10 张 9 中和 10 中的盲评出口门验收）。参考集由你确认，生成产物不会自动混进去。 Q: 伴侣会知道我这本书读到哪了吗？ A: 不会，这是我们明确没做的一步。伴侣的记忆来自你们的对话和你给的记录，不追踪你的阅读进度——「TA 记得你读到第几章」听起来浪漫，实际是把阅读行为变成了监控。这条线我们不跨。 Q: 把伴侣导出成角色卡送人，对方能看到我们的聊天记录吗？ A: 不能。导出时逐条勾选哪些记忆随卡带走，默认一条都不选；被勾选的记忆以世界书条目形式进卡，聊天记录本身永远不出你的手机。 ### 315 本 txt 一次全选导入：一台真机的阅读器性能账本 URL: https://foreverse.app/zh/blog/315-books-in-8-seconds Published: 2026-07-11 · Author: Deng Binjie 拿一台 vivo 折叠屏真机灌入真实手机存储快照，把扫出的 315 本网文 txt 一次性全选导入，批量入库约 8 秒；随机进 10 本书全部正常打开，7 万多帧渲染统计卡顿率 0.49%。这篇公开完整测法和两台真机的数字，包括我们在测试里修掉的一个数据安全隐患。 Q: 批量导入几百本书需要多久？ A: 实测 315 本真实 txt（合计约 1GB）一次全选，批量入库约 8 秒，全程无卡死无崩溃。另一台小米真机导入 82 本约 7 秒。导入是「先登记后转换」：书先以待转换状态上架，第一次打开时才做全文解析。 Q: 「待转换」的书第一次打开要等多久？ A: vivo 真机上实测 3.3 到 6.8 秒（含全文章节解析和首屏排版），小米上 9.5 到 14.5 秒。之后打开跳过全文转换走索引，实测小米上已转换书 5 秒左右，具体随设备和书籍体积变化。 Q: 卡顿率 0.49% 是怎么测出来的？ A: 用 Android 系统自带的帧渲染统计（gfxinfo）：在真机上完成整轮导入加随机进书操作后读取累计数据，71320 帧里卡顿帧占 0.49%，99 分位帧耗时 31ms。这是系统口径的数字，不是自报的。 Q: 从旧阅读器搬家，隐藏目录里的书能扫到吗？ A: 能。智能扫描覆盖了常见阅读器的存书目录，包括浏览器下载目录和以点开头的隐藏文件夹；已授权过的目录直接扫，扫描在 315 本规模下秒级出结果。 ### AI 怎么知道这句话是谁说的？说话人归因的五个坑 URL: https://foreverse.app/zh/blog/who-said-that Published: 2026-07-11 · Updated: 2026-07-14 · Author: Deng Binjie 把小说演成视觉小说，第一关是给每句引语找到说话人。这篇拆解说话人归因的五个真实案例：后置说话从句、受话人陷阱、交替对话的边界、句号归属，以及为什么我们最终选择「宁可放旁白，不冒错认的险」。含真书人工标注的核对结果。 Q: 为什么不直接让大模型判断谁在说话？ A: 成本和确定性。一本 100 章的书按段落调用模型判断说话人，费用会转嫁给用户，而且模型输出不稳定，同一段落两次判断可能给出不同答案。规则方案零成本、离线可用、结果可复现，代价是覆盖率低——我们接受这个代价。 Q: 规则方案的准确率到底怎么验证的？ A: 人工标注。我们从一本三百多万字的已出版网文里摘出 15 段对话密集的段落，逐句人工标注说话人作为标准答案，再跑归因规则对比。强归因部分的精确率是 100%——15 段里没有一处错认；拿不准的句子全部按设计进了旁白。 Q: 归因失败的句子会怎么呈现？ A: 以旁白形式演出，引语原文保留。观感上是「无名牌的台词」，读者靠上下文自己判断——这正是纸书读者一直在做的事，不损失任何信息。 Q: 以后会提高归因覆盖率吗？ A: 会，但底线不变：新增规则必须先过真书标注核对，错误归因零容忍。覆盖率可以慢慢爬，把 A 的话安到 B 头上这种事故一次都不能有。 ### 把书架上的小说一键开成文字游戏：剧场模式实测记录 URL: https://foreverse.app/zh/blog/novel-to-visual-novel Published: 2026-07-11 · Updated: 2026-07-14 · Author: Deng Binjie 把正在读的小说一键切成视觉小说演出：逐字台词、说话人名牌、AI 场景背景、剧情选择肢。这篇是剧场模式在两本书上的完整实测记录——包括它怎么做到不转换、不丢进度、选出来的剧情真实写回书里，以及我们在「谁在说话」这个问题上交过的学费。 Q: 剧场模式需要把小说转换成游戏文件吗？ A: 不需要。剧场直接读这本书当前分支的正文，实时编排成演出，不产生新工程、不迁移数据。退出剧场回到阅读器，进度就停在剧场演到那段所在的页；重进剧场，从上次的位置接着演。 Q: 把正文编排成舞台脚本这一步要花多少钱？ A: 零。编排完全由本地规则完成，不调用任何 AI 模型，离线也能进剧场。只有两件事按你的选择走模型：点「生成此景」出背景图（生成前弹费用确认），以及确认剧情选择肢之后的 AI 续写。 Q: 在剧场里选的剧情，会改动我的原书吗？ A: 以分支形式写入，原文永远保留。二击确认一个走向后，AI 续写的新剧情落成这本书的一条正式分支——回阅读器可以接着读，重进剧场可以接着演，不想要就切回原文线。 Q: 什么书开剧场效果最好？ A: 对话密度高的书：都市、言情、群像、轻小说。散文向的书也能开，只是更多句子会以旁白形式演出。没有配置过角色的书同样能进剧场，全部内容按旁白演。 ### Writing Fanfiction With AI: What Actually Works, From a Reader's Request URL: https://foreverse.app/blog/ai-fanfiction-honestly Published: 2026-07-07 · Author: Deng Binjie Someone on our waitlist asked for TGCF-style fanfiction support. An honest assessment: where AI fanfic quality actually stands, how to feed canon so characters stay in character, why branching solves the three-endings problem, and the problems (voice drift, OOC, platform disclosure rules) that no tool has solved. Q: How good is AI at writing fanfiction right now? A: Short pieces (1,000-3,000 words) are genuinely publishable: scene work and dialogue pacing hold up. Long fics are where it breaks — past 20,000 words, characterization and planted details drift unless you support them structurally. The fix is not a better model; it is character dossiers and pinned worldbook entries injected on every request. Q: How do I keep AI from writing characters OOC? A: In order of measured effectiveness: example dialogue (3-5 canon exchanges that carry the character's voice beat any adjective list), negative constraints ('he never apologizes first' prevents more drift than ten positive traits), and generating multiple candidates for key scenes so you judge in-character-ness yourself. The OOC verdict always belongs to you; tools just put candidates side by side. Q: I want three different endings from the same scene. How do I manage that? A: That is exactly what branching is for: fork the story at the divergence point and grow the fix-it, the tragedy, and the what-if line in parallel. Each branch keeps its own full text and context. In single-document tools you juggle three files; in a branch tree they hang off the same origin and you switch between them freely. Q: What about copyright and posting rules for AI-assisted fanfic? A: Fanwork copyright status varies by fandom and platform, and AI involvement does not change that baseline — but most archives (AO3 included) now have disclosure expectations for AI-assisted work, and hiding it gets works reported. Post only where permitted, disclose honestly, never commercialize. Not legal advice; check your archive's current policy. ### 用 AI 写同人文靠谱吗？回答一位想续写耽美同人的等待名单用户 URL: https://foreverse.app/zh/blog/write-fanfic-with-ai Published: 2026-07-07 · Author: Deng Binjie 等待名单里一位用户留言想用 AI 写 TGCF 式的耽美同人。这篇认真回答：AI 写同人的真实水平在哪、原作人设怎么喂给模型才不崩、分支怎么解决「一个梗想写三个走向」、以及哪些坑（OOC、文风飘、审查尺度）现阶段绕不开。 Q: AI 现在写同人文是什么水平？ A: 旁白叙述和场景描写已经可用，短打（1000-3000 字）质量稳定；长篇的坑在一致性——写到几万字后人设和伏笔会漂。补救靠结构而不是靠模型：把人设做成角色档案、把设定做成世界书常驻条目，每轮都稳定注入，漂移会显著减少。 Q: 怎么让 AI 不 OOC（不写崩人设）？ A: 三件事按重要性排：示例对话（从原作摘 3-5 段最能代表角色语气的对话，比任何形容词都管用）、把「这个角色绝不会做什么」写进档案（负面约束比正面描述更防崩）、重要情节让模型一次给多个候选再由你挑。判断 OOC 的始终是你，工具能做的是把候选摆到一起。 Q: 同一个梗想写好几个走向，怎么管理？ A: 这正是分支续写的场景：从同一个分歧点开多条分支，HE 一条、BE 一条、if 线一条，每条独立生长互不覆盖。单文档工具里你只能存三个文件手动对照；分支树里它们并排挂在同一个原点上，随时切换对比。 Q: AI 写同人的版权和发布要注意什么？ A: 同人创作本身的版权状态因作品和平台而异，AI 参与不改变这一点，但多数发布平台（AO3、LOFTER 等）要求标注 AI 参与程度，隐瞒会被举报。建议：只在允许的分区发布、如实标注、商用零容忍。这不是法律建议，拿不准就查目标平台的最新政策。 ### AI Token 真的越来越便宜吗？看完定价数据后我改了自己的用法 URL: https://foreverse.app/zh/blog/are-ai-tokens-getting-cheaper Published: 2026-07-06 · Updated: 2026-07-14 · Author: Deng Binjie a16z 把 LLM 推理降价称作「LLMflation」：同等能力的 token 价格约每年降 10 倍；Epoch AI 的口径甚至更激进。但旗舰模型定价三年没动，推理模型的思考 token 让单次任务更贵。一篇买家视角的定价观察，附我们自己改用法的过程。 Q: 「同等能力降价 10 倍」是怎么算出来的？ A: a16z 的 LLMflation 分析取「在固定基准上达到同一分数的最便宜模型」作比较：达到 GPT-3 水平（MMLU 约 42 分）的成本从 2021 年的每百万 token 60 美元跌到 2024 年的 0.06 美元，三年千倍，折合每年约 10 倍。Epoch AI 对多个基准的独立统计给出每年 9 到 900 倍不等的下降区间，取决于任务难度。 Q: 那为什么我的 AI 账单没有降？ A: 三个原因：一，旗舰模型定价基本不随时间下降，降价发生在「旧能力被新的便宜模型复刻」；二，推理模型会产出大量思考 token，单次任务消耗反而上升；三，用量本身在涨——agent 类应用一次任务几十次调用。降价红利属于愿意换模型的人，不属于始终买最贵的人。 Q: 现在（2026 年中）主流模型的价格量级是多少？ A: 以各家官方定价页为准的量级感受：DeepSeek 的 flash 级模型输入价在每百万 token 零点几美元；OpenAI / Anthropic / Google 的旗舰在每百万 token 几美元到十几美元；各家 mini / flash 档普遍比自家旗舰便宜 10 到 25 倍。加上缓存折扣，实际有效价格还能再降一半以上。 Q: 普通用户怎么吃到降价红利？ A: 两条：按场景分配模型——日常续写和闲聊用 flash 档，关键章节和复杂推演才上旗舰，大多数场景 flash 档已经够用；以及用 BYOK 保持随时可换——模型市场一年洗一次牌，绑死在单一订阅上就等于放弃了每年 10 倍的降价曲线。 ### AI Token 缓存为什么越来越重要？一次 60 轮续写实测的账单笔记 URL: https://foreverse.app/zh/blog/why-prompt-caching-matters Published: 2026-07-05 · Updated: 2026-07-07 · Author: Deng Binjie Prompt 缓存（KV Cache）已成为大模型计费的隐形折扣：DeepSeek 缓存命中价约为原价 1/10，Anthropic 缓存读取按 10% 计费，OpenAI 自动打五折。这篇是一份实测笔记：60 轮连续续写的真实账单、命中率从 91% 掉到 0% 的事故现场，和四家供应商的规则差异。 Q: Prompt 缓存到底缓存了什么？ A: 缓存的是模型对「已经见过的前缀」的中间计算结果（KV Cache）。同一段上下文第二次出现时，供应商不必重新计算这部分注意力状态，直接从缓存加载，所以能以远低于原价的价格计费。缓存按前缀匹配：只要请求开头和上次完全一致，一致的部分就能命中。 Q: 各家的缓存折扣差多少？ A: 以官方定价页为准的量级：DeepSeek 缓存命中的输入约为未命中价的 1/10；Anthropic 缓存读取按基础输入价的 10% 计费（写入缓存加收 25%）；OpenAI 对 1024 token 以上的重复前缀自动打对折；Gemini 提供显式与隐式上下文缓存，命中部分同样按折扣价计。 Q: 为什么长对话反而越聊越便宜？ A: 对话第 N 轮的输入 = 前 N-1 轮的全部历史 + 新一句。历史部分正是上一轮刚出现过的前缀，天然可命中缓存。轮次越多，可缓存的占比越高——第 50 轮时输入的 95% 以上都是折扣价 token，边际成本主要来自新增文字和输出。 Q: 什么操作会把缓存搞失效？ A: 任何改动前缀的行为：往系统提示里插入当前时间戳、每轮重排世界书条目顺序、在历史中间编辑一条旧消息。缓存按前缀严格匹配，前缀变一个字，之后的全部重新计费。工具设计上应把「稳定的内容放前面、易变的内容放后面」。 ### Prompt Caching Explained: What 60 Continuation Requests Taught Us About LLM Bills URL: https://foreverse.app/blog/prompt-caching-explained Published: 2026-07-04 · Updated: 2026-07-07 · Author: Deng Binjie Prompt caching is the largest line-item discount in LLM billing: DeepSeek bills cache hits at roughly one-tenth of the miss price, Anthropic reads cost 10% of base input, OpenAI halves repeated prefixes automatically. Notes from a 60-request field test, including the timestamp bug that dropped our hit rate from 91% to zero. Q: What exactly does prompt caching cache? A: The model's intermediate attention state (the KV cache) for a prefix it has already processed. When a request starts with the exact same content as a previous one, the provider loads that state instead of recomputing it, and bills the matched portion at a steep discount. Matching is strictly prefix-based: identical from the first character up to the divergence point. Q: How big are the discounts across providers? A: Order-of-magnitude, per official pricing pages: DeepSeek bills cache-hit input at roughly one-tenth of the miss price; Anthropic cache reads cost 10% of base input (writes carry a 25% premium); OpenAI automatically halves input for repeated prefixes over 1,024 tokens; Gemini offers implicit and explicit context caching with discounted hit pricing. Q: Why do long conversations get cheaper per turn? A: Turn N's input is the full history of turns 1..N-1 plus one new message. That history is precisely the prefix the provider just processed, so it hits cache naturally. The deeper the conversation, the higher the cacheable share. By turn 50, over 95% of input tokens are discount-priced, and marginal cost comes almost entirely from new text and output. Q: What breaks caching? A: Anything that mutates the prefix: injecting a per-request timestamp into the system prompt, reordering lorebook entries each turn, or editing an old message mid-history. Matching is exact. Change one character early in the prompt and everything after it bills at full price. Design rule: stable content first, volatile content last. ### C.AI 又改规则了？我们陪一位用户把养了八个月的角色搬了家 URL: https://foreverse.app/zh/blog/cai-refugee-migration-guide Published: 2026-07-03 · Updated: 2026-07-14 · Author: Deng Binjie Character.AI 类平台的老问题：过滤收紧、角色下架、记忆变差、数据带不走。这篇是一次真实迁移的过程记录：怎么把平台上的角色重建成 chara_card_v3 角色卡、几千条聊天记录怎么蒸馏成关系档案、BYOK 换模型，以及搬完之后哪些地方不如从前。 Q: C.AI 的角色能直接导出成角色卡吗？ A: 官方没有导出按钮，角色定义（尤其别人创建的角色）拿不到原文。可行的路径是重建：把角色的公开简介、你记忆里的性格要点、加上从聊天记录里摘出的经典对话，按 chara_card_v3 格式重写一张卡。社区也有第三方抓取工具，但依赖非官方接口，随时可能失效，重建反而更可控。 Q: 聊了几千条的记忆怎么办？ A: 全量搬运没有意义——模型每轮能看的上下文有限，几千条原始记录塞不进也不需要。正确做法是蒸馏：把关系的关键节点（怎么认识的、几次重要冲突与和解、彼此的称呼和梗）整理成一两千字的「关系简历」，写进角色卡的场景设定或世界书常驻条目。新平台上的第一句话，就能接住八个月的感情线。 Q: 换到 BYOK 应用后聊天质量会下降吗？ A: 取决于你选的模型，上限通常更高：C.AI 用的是它自家的专有模型，你没得选；BYOK 下 DeepSeek、Gemini、Claude 随便换，角色扮演口碑好的模型可以按卡分配。落差主要在两点：需要自己配 Key（一次性成本），以及没有平台内置的社区推荐流——角色要自己找、自己养。 Q: 怎么避免下一次「被迁移」？ A: 认准三个可迁移标志：角色以开放格式存储（chara_card_v3 的 PNG/JSON 文件，而不是平台数据库里的一行记录）；聊天记录可以一键导出成通用格式；模型接入走你自己的 Key。三条都满足时，换应用只是换个壳，你的角色、记忆、模型关系全部原地带走。 ### Are AI Tokens Getting Cheaper? The Pricing Data Says Yes and Your Bill Says No URL: https://foreverse.app/blog/llm-token-price-trends Published: 2026-07-02 · Updated: 2026-07-14 · Author: Deng Binjie a16z calls it LLMflation: the price of a fixed level of LLM capability drops roughly 10x per year, and Epoch AI measures 9-900x depending on the task. Meanwhile frontier list prices have been flat for three years and reasoning tokens inflate per-task cost. A buyer's memo, with the changes we made to our own usage. Q: Where does the “10x cheaper per year” figure come from? A: a16z's LLMflation analysis tracks the cheapest model that reaches a fixed benchmark score: hitting GPT-3-level MMLU (~42) cost $60 per million tokens in 2021 and $0.06 by late 2024, a 1,000x drop in three years, roughly 10x per year. Epoch AI's independent study across multiple benchmarks finds annual declines between 9x and 900x depending on task difficulty, with ~50x for the median. Q: Then why doesn't my AI bill go down? A: Three reasons. Frontier list prices barely fall; the decline happens when new cheap models replicate old capabilities. Reasoning models emit large volumes of billed thinking tokens, raising per-task consumption 5-20x. And usage itself grows; agent workloads fire dozens of calls per task. The dividend goes to people willing to switch models, not to people who always buy the most expensive one. Q: What do prices look like in mid-2026? A: Order-of-magnitude, per official pricing pages: flash-tier models (DeepSeek, Gemini Flash class) charge tenths of a dollar per million input tokens; frontier models from OpenAI, Anthropic and Google sit at a few to tens of dollars per million; every vendor's mini/flash tier runs 10-25x cheaper than its own flagship. Prompt-cache discounts cut effective prices by half or more on top. Q: How does a regular user capture the decline? A: Tier your scenarios: run everyday continuation and casual chat on flash-tier models and reserve the flagship for pivotal chapters and complex reasoning. Flash tiers are already sufficient for most fiction work. And stay switchable with BYOK: the model market reshuffles yearly, and locking into a single subscription forfeits the 10x/year curve entirely. ### 世界书塞了 237 条设定，AI 还是叫错名字：一次完整的排障记录 URL: https://foreverse.app/zh/blog/lorebook-writing-guide Published: 2026-07-01 · Updated: 2026-07-14 · Author: Deng Binjie 一位创作者的世界书写了 237 条，AI 第三轮就把师父认成仇人。这篇是完整排障记录：从关键词扫描、别名覆盖、token 预算到递归触发，一步步找出四种「无声失效」，顺带得出几条可复用的世界书写法。附症状排查表。 Q: 世界书和角色卡的 description 有什么区别？ A: description 是常驻的：每一轮对话都完整占用上下文，适合放角色的核心人设。世界书默认是按需的：只有对话里出现关键词才注入对应条目，适合放数量大、但不是每轮都用得上的设定——地名、支线人物、历史事件。把所有设定都塞进 description，等于每轮都在为用不上的信息付 token。 Q: 世界书条目没触发，怎么排查？ A: 按顺序查四点：一，关键词是否真的出现在最近几轮对话里（触发靠扫描最近消息，不是全文）；二，别名有没有覆盖——对话里叫「老李」，关键词只写了「李云山」就不会触发；三，扫描深度是否太浅，关键词出现在更早的楼层就扫不到；四，token 预算是否被前面的条目吃满，后面的条目被静默丢弃。 Q: 一本世界书多大算太大？ A: 条目数量本身不是问题，几百条也没关系——问题在单轮注入量。触发式条目每轮实际注入的总量受 token 预算限制，常驻（constant）条目则每轮必进。经验值：常驻条目控制在 5 条以内，单条内容 200 字以内；触发条目单条不超过 300 字。写满 2000 字的「王国编年史」不如拆成 10 条各管一段。 Q: V2 和 V3 的世界书有什么区别？ A: V3（chara_card_v3）在条目上增加了 decorators——以 @@ 开头的指令，可以精确控制注入深度、触发条件等行为，世界书的表达力更强。V2 卡的世界书在支持 V3 的应用里可以正常使用，反过来 V3 特有的 decorators 在旧应用里会被忽略。导入前确认应用支持的规范版本，能省掉很多「为什么行为不一致」的困惑。 ### Turn Any EPUB Into an AI Audiobook: Three Weeks of Commute Testing, Honestly Reported URL: https://foreverse.app/blog/epub-to-ai-audiobook Published: 2026-06-29 · Updated: 2026-07-07 · Author: Deng Binjie Most books never get an audiobook; studio narration costs thousands of dollars per title. Neural TTS in 2026 is good enough to fix that, so we spent three weeks of commutes listening to a 900-chapter webnovel and wrote down what broke: asterisks read aloud, chapter-boundary stutters, split progress, and what each cost to fix. Q: How good is AI narration compared to a human audiobook in 2026? A: For single-narrator prose it is close enough that most listeners stop noticing after a chapter. In our blind test, three out of five colleagues could not tell a neural TTS voice from a human narrator on a 2,000-word passage. The remaining gap is multi-character dialogue: human narrators switch voices per character; most TTS pipelines still read everything in one voice. Q: What does it cost to listen to a whole novel with TTS? A: TTS is billed per character. A 3,000-word chapter runs roughly 16,000-18,000 characters; at typical provider rates that is a fraction of a cent to a few cents per chapter depending on the voice tier. A full 300-chapter webnovel usually lands in the single-digit dollars, with your own API key, at provider list price, no subscription on top. Q: Why does my TTS app read out asterisks and chapter dividers? A: Because it sent the raw text to the speech engine. EPUB and TXT files from the wild are full of things that should never be spoken: scene-break dashes, translator notes, promo lines, emoji. An app that cares about listening strips these before synthesis. If you hear a warm voice solemnly pronounce three asterisks, the app skipped its homework. Q: Can reading and listening share one progress? A: They should, and this is the feature that decides whether TTS is a gimmick or a habit. You listen through chapter 230 on the commute, open the book at home, and the page picks up where your ears left off. If the app keeps two separate positions, you will spend every evening scrolling to find yourself. ### Lorebook Not Triggering? Three Support Tickets and What Fixed Each One URL: https://foreverse.app/blog/lorebook-world-info-not-triggering Published: 2026-06-24 · Updated: 2026-07-07 · Author: Deng Binjie Three real cases of world info failing without an error message: keywords that never matched, a token budget quietly eaten by one 800-word entry, and a scan window the conversation outran. How World Info injection actually works, what fixed each ticket, plus a symptom table. Q: How does a lorebook decide which entries to inject? A: On every message, the app scans the most recent turns of the conversation for each entry's keywords. Matching entries are activated and inserted into the prompt at their configured position, in priority order, until the lorebook's token budget is spent. Entries beyond the budget are dropped silently; the model never knows they existed. Q: Why does my entry trigger sometimes but not always? A: Almost always scan depth. Keywords are matched against the last N messages, not the whole chat. If the dragon was last mentioned six turns ago and scan depth is four, the entry sleeps. Raise the depth, add the keyword to related entries so recursion keeps it warm, or make the entry constant if it genuinely matters every turn. Q: Is a big lorebook bad for quality? A: Entry count is fine; per-turn injection volume is what hurts. Constant entries cost context every single turn, so keep them under five, each under ~150 words. Triggered entries compete for one token budget, and a single 800-word essay entry can starve three concise ones. Split epics into one-fact-per-entry cards. Q: Do V2 lorebooks work in V3 apps? A: Yes. chara_card_v3 is backward compatible: V2 world info imports and runs unchanged. V3 adds decorators, @@-prefixed directives on entries that control injection depth and activation conditions. Those extras are ignored by V2-only apps, which is the main source of 'same card, different behavior' reports. ### 角色卡导入失败的 9 个原因：PNG、JSON、世界书一次排查清楚 URL: https://foreverse.app/zh/blog/character-card-import-troubleshooting Published: 2026-06-05 · Author: Deng Binjie 角色卡导入失败 90% 是这 9 个原因：微信压缩抹掉 PNG 隐写、JSON 编码损坏、v1 老卡、改后缀、重名冲突等。附手机端特有的坑和一套 30 秒自检流程。 Q: 为什么微信传过来的角色卡导不进去？ A: 微信和 QQ 默认压缩图片，压缩过程会重新编码 PNG，把藏在 tEXt 块里的角色数据整块抹掉。让对方以「文件」方式发送原图，或者打包成 zip 再传，就能保住卡片数据。 Q: 怎么快速判断一张 PNG 卡里有没有数据？ A: 看文件大小。一张正常的角色卡 PNG 通常在 100KB 到 3MB 之间；如果一张「精美立绘」只有几十 KB，多半已经被某个传输环节重新压缩过，卡片数据凶多吉少。最稳的办法是导入一次试试，或用文本编辑器打开搜索 chara 关键字。 Q: v1 老卡还能用吗？ A: 大部分现代客户端（包括 SillyTavern 和 Foreverse）仍然向下兼容 v1 卡的基础字段，但 v1 没有备选问候语、角色书等结构，体验会缺一块。建议用编辑器把老卡升级到 v3 再用。 Q: 角色卡带的世界书为什么没生效？ A: 看两个地方：一是导入时世界书是否被一并导入（部分客户端需要手动确认），二是世界书是绑定到角色还是全局启用。绑定角色的世界书只在和该角色聊天时注入，全局的才会处处生效。 ### How to Run SillyTavern on Android or iPhone in 2026: Termux, Cloud, or Native App URL: https://foreverse.app/blog/run-sillytavern-on-android-iphone Published: 2026-06-03 · Updated: 2026-07-19 · Author: Deng Binjie Four real ways to get SillyTavern on a phone in 2026 — Termux on Android, a cloud VPS, LAN access to your PC, or a native tavern-style app. Costs, trade-offs, and what survives the migration. Q: Can SillyTavern run natively on an iPhone? A: No. SillyTavern is a Node.js server application, and iOS does not allow long-running server processes. iPhone users have two realistic options: open a cloud- or PC-hosted instance in Safari, or use a natively built tavern-style app that reads the same character cards. Q: Why does my Termux SillyTavern stop when the screen locks? A: Android freezes background processes to reclaim memory, and a Node.js server inside Termux is a prime target. termux-wake-lock and disabling battery optimization help, but aggressive vendor ROMs often override both. It is a platform behavior, not a SillyTavern bug. Q: Will my character cards work across all four routes? A: Yes — PNG and JSON character cards plus lorebook JSON files are portable everywhere. Chat history is stored as jsonl inside the SillyTavern ecosystem; native apps vary, so check that the app explicitly imports SillyTavern data before you commit. Q: Is a shared cloud tavern safe to use? A: Only as safe as the person running it. On a shared instance, your API keys, cards, and chat logs all pass through someone else's server. Self-host if you can; if you cannot, treat any service that asks for your API key with appropriate suspicion. ### 2026 年手机玩 SillyTavern 酒馆的四条路：Termux、云端、局域网、原生 App URL: https://foreverse.app/zh/blog/sillytavern-on-phone-2026 Published: 2026-06-02 · Updated: 2026-07-14 · Author: Deng Binjie 想在安卓或 iPhone 上玩 SillyTavern？这篇把 Termux 本地部署、云端酒馆、局域网访问、原生酒馆 App 四条路线的真实体验、成本和坑一次讲清，附角色卡迁移建议。 Q: Termux 跑酒馆，锁屏后为什么会断？ A: Android 会冻结后台进程回收内存，Termux 里的 Node.js 服务首当其冲。可以用 termux-wake-lock 和关闭电池优化来缓解，但厂商 ROM（尤其国产定制系统）的激进省电策略经常绕过这些设置。这是系统机制，不是酒馆的 bug。 Q: iPhone 能本地跑 SillyTavern 吗？ A: 不能。SillyTavern 是 Node.js 应用，iOS 不允许运行这类常驻服务进程。iPhone 用户只有两条路：浏览器访问部署在云端或家里电脑上的酒馆，或者使用原生开发的酒馆类 App。 Q: 换路线后，原来的角色卡和聊天记录能带走吗？ A: 角色卡（PNG / JSON）和世界书是通用文件，四条路线之间都能互相导入。聊天记录在 SillyTavern 体系内是 jsonl 文件，可以备份迁移；迁去原生 App 则取决于对方是否兼容 SillyTavern 的数据格式，导入前先确认。 Q: 云端酒馆安全吗？ A: 取决于谁在运维。自己买 VPS 自己部署，数据归你管；用别人搭好的「公共酒馆」，你的 API Key、角色卡和聊天记录都过他的服务器，这层信任要自己评估。任何要求你把 Key 填进陌生网页的服务都值得多想三秒。 ### Why Your AI Roleplay Partner Forgets Everything — and What Actually Fixes It URL: https://foreverse.app/blog/why-ai-roleplay-chats-forget Published: 2026-05-27 · Author: Deng Binjie AI roleplay memory loss isn't a bug, it's architecture: context windows, summary compression, and platform-owned history. What lorebooks, character cards, and on-device chat libraries actually fix — and what they can't. Q: Why do AI characters forget things from earlier in the chat? A: Every model reads a fixed budget of text per reply — the context window. Once a chat outgrows it, something must be dropped, and naive clients silently drop your oldest messages. The model never 'knew' your whole story; it only ever saw the slice that fit. Q: Does a bigger context window solve roleplay memory? A: It delays the problem and raises the bill, but does not solve it. A 200k-token window still fills after weeks of play, costs scale with every token you resend, and models demonstrably pay less attention to the middle of very long contexts. Structure beats brute force. Q: What is a lorebook and how does it help memory? A: A lorebook is a set of keyed entries — facts about people, places, and events — that get injected into the prompt only when relevant keywords appear in the recent conversation. It acts as cheap, targeted long-term memory: the dragon's death costs zero tokens until someone mentions the dragon. Q: Can I keep my chat history if a platform shuts down? A: Only if your history lives somewhere you control. Platform-hosted services have deleted chats with policy changes or shutdowns before. Clients that store chats on your device in open formats (like SillyTavern's jsonl) make your story archive independent of any company's roadmap. ### AI 续写小说为什么总把书写崩？我们换了一种结构来解决 URL: https://foreverse.app/zh/blog/ai-continue-novel-branches Published: 2026-05-26 · Author: Deng Binjie AI 续写小说最大的痛不是文笔是结构：上下文丢失、风格漂移、改了就回不去。这篇讲分支续写的思路——像管理代码版本一样管理剧情线，烂尾文自救的工程化方案。 Q: AI 续写为什么会忘记前文设定？ A: 模型一次能看到的文字量（上下文窗口）是有限的，几十万字的书不可能整本塞进去。朴素做法只送最近几千字，前面埋的伏笔自然全丢。解法是给书建立结构化的设定档案，续写时按相关性挑选注入，而不是只看「最近的文字」。 Q: 分支续写和普通的「重新生成」有什么区别？ A: 重新生成是覆盖：新结果顶掉旧结果，旧的就没了。分支是并存：每个方向都保留完整的文字和上下文，可以随时回头、对比、继续往深里写。前者像抽卡，后者像创作。 Q: 续写能模仿原作者的文风吗？ A: 能接近，做不到完美。靠谱的做法是让模型从原文中学习用词、节奏和对话密度，并允许你按章节微调提示词。任何宣称「百分百还原文风」的产品都在夸大。 Q: 烂尾的连载文能用 AI 救回来吗？ A: 可以试，而且这是分支续写最合适的场景之一：从烂尾点开一条分支，把你心中「本该有的结局」写出来；不满意就再开一条。原文永远完好，你失去的只是一点电费。 ### Character Card Formats Explained: v1, v2, and chara_card_v3 — What Actually Imports URL: https://foreverse.app/blog/character-card-formats-explained Published: 2026-05-20 · Author: Deng Binjie How character cards really work: data hidden in PNG tEXt chunks, what each spec generation added (alternate greetings, character books, v3 assets), why cards break in transit, and a distribution checklist for creators. Q: How is character data stored inside a PNG? A: As base64-encoded JSON inside the PNG's tEXt metadata chunks — 'chara' for v2 data and 'ccv3' for v3. The image you see is just the portrait; the character lives in metadata. Any pipeline that re-encodes the image (messenger compression, photo-library 'optimization') strips those chunks and kills the card. Q: What did v2 add over v1? A: v1 was six flat fields: name, description, personality, scenario, first message, and example dialogue. v2 wrapped them in a versioned envelope and added the fields the community had been hacking in: alternate greetings, system prompt overrides, creator notes, tags, and most importantly an embeddable character book (mini-lorebook). Q: Is chara_card_v3 backwards compatible? A: Largely yes by design — v3 keeps the v2 envelope shape, so a v3 card usually degrades gracefully in a v2-only client (new fields like assets, multilingual creator notes, and decorators are simply ignored). A v2 card in a v3 client just works. Q: Should creators distribute PNG or JSON? A: Both. PNG for collectors and galleries — it carries the portrait. JSON for reliability — it survives chat apps, email, and re-uploads that would strip PNG metadata. Label the spec version and note whether a character book is embedded; it preempts most 'card won't import' reports. ### BYOK 不是中转：在手机上自带 API Key 的正确姿势 URL: https://foreverse.app/zh/blog/byok-on-mobile Published: 2026-05-19 · Author: Deng Binjie 买中转 Key 被跑路、填进网页被盗刷——这篇讲清 BYOK（自带 API Key）的判别标准：密钥是否离开设备、请求是否直连官方、价格是否透明，以及手机端管理多家供应商密钥的实践。 Q: BYOK 和 API 中转有什么区别？ A: BYOK 是你直接持有官方（或你信任的供应商）的 Key，应用拿它直连该供应商的服务器；中转是把别人的 Key 二次销售给你，你的请求全部过中转商的服务器。前者的信任链是「你—供应商」两点，后者多了一个可以看到你全部对话、随时可能跑路的中间人。 Q: 怎么验证一个 App 是不是真 BYOK？ A: 看三点：一，填 Key 时是否允许你自定义 API 地址（真 BYOK 必然支持）；二，断开它家账号体系后 BYOK 功能是否照常工作；三，有条件的话抓包看请求是否直达供应商域名。只要 Key 需要「上传到我们的服务器托管」，就不是本文说的 BYOK。 Q: 手机丢了，存在里面的 Key 会泄露吗？ A: 取决于应用怎么存。合格的做法是用系统级加密（Android Keystore 这类硬件背书的方案）落盘，应用数据导出后没有密钥也解不开。再加一层保险：给重要 Key 设消费上限，丢失后第一时间在供应商后台吊销。 Q: BYOK 会比官方 App 便宜吗？ A: 对重度用户通常是。你按供应商的原始价格付费，没有订阅费和中间加价；轻度用户则可能不如包月划算。BYOK 的核心收益其实不是省钱，是选择权——60 多家供应商、上百个模型，按场景挑最合适的。 ### BYOK, Actually: How to Tell If an App Really Keeps Your API Key on Device URL: https://foreverse.app/blog/what-byok-actually-means Published: 2026-05-13 · Author: Deng Binjie BYOK has become a checkbox word. The real definition: your key stays on your device and requests go straight to the provider. Here are the red flags, the three-step verification, and how key storage should work on a phone. Q: What does BYOK mean in AI apps? A: Bring Your Own Key: you create an API key with a provider (OpenAI, Anthropic, DeepSeek, etc.), the app stores it on your device, and every request goes from your device directly to that provider. You pay raw provider prices. If the key or your conversations route through the app vendor's servers, it is proxy billing wearing a BYOK costume. Q: How do I verify an app is genuinely BYOK? A: Three tests: (1) it lets you set a custom API base URL, not just paste a key; (2) BYOK chats keep working when you log out of the app's own account system; (3) if you can, watch the traffic — requests should hit the provider's domain, not the vendor's. Any app that 'syncs your key to our cloud' fails the definition. Q: Where should a mobile app store API keys? A: In hardware-backed encrypted storage (Android Keystore / iOS Keychain), excluded from cloud backups, never written to logs or analytics. A copied app-data folder without the device's hardware key should be useless to an attacker. Q: Is BYOK cheaper than a subscription? A: For heavy users, usually — you pay provider list price with no middleman margin. For light users a flat subscription can win. The durable benefit is not price but leverage: when one provider raises prices or degrades a model, you switch endpoints in a minute instead of migrating your life. ### 通勤路上「听」完 300 万字网文：AI 朗读的实测笔记 URL: https://foreverse.app/zh/blog/listen-to-webnovels-tts Published: 2026-05-12 · Updated: 2026-07-14 · Author: Deng Binjie AI 朗读听网文的真实体验：现代 TTS 和十年前机械音的差距、听书最烦的章节边界和错别字问题怎么处理、流量和缓存的取舍，以及挑选音色的实用建议。 Q: AI 朗读和真人有声书差距还大吗？ A: 旁白叙述部分已经非常接近，长时间听不出疲惫感；差距主要在多角色对手戏——真人演播会为每个角色切换声线，AI 朗读目前多数还是单音色到底。听爽文、种田文几乎无感，听群像戏会想念真人。 Q: 听书很费流量吗？ A: 语音合成按字数计，一章 3000 字大约对应几百 KB 到 2MB 音频（取决于音质档位）。Wi-Fi 下提前缓存整卷，地铁里零流量播放，是目前最省的方案。 Q: 网文里的错别字和怪符号会被读出来吗？ A: 处理得好的应用会在送往语音引擎前清洗文本：剥掉分割线、广告尾巴、表情符号，多数错别字模型会按上下文读出正确发音。生僻人名偶尔翻车，听两章就习惯了。 Q: 可以边听边看吗？ A: 可以，这其实是被低估的用法：眼睛跟着高亮走，耳朵听声音，注意力比单独看或单独听都集中。午休躺着听、通勤站着看，一本书两种姿势无缝接力。 ### Rerolling Is Gambling. Branching Is Writing. A Better Way to Use AI on Long Fiction URL: https://foreverse.app/blog/branching-beats-rerolling Published: 2026-05-06 · Author: Deng Binjie Why the reroll button quietly ruins AI-assisted fiction: destructive generation, lost alternatives, sunk-cost plotting. Branching — version control for narrative — keeps every draft alive and makes long stories coherent. Q: What is branching in AI fiction? A: Instead of regenerating over a passage you dislike, you fork the story at any paragraph and grow an alternative line alongside the original. Every branch keeps its own full text and context. It is version control applied to narrative: nothing is overwritten, everything is comparable. Q: Why is rerolling worse than branching? A: Rerolling is destructive — each new roll replaces the last, so a great take you rolled past is gone forever. It also trains writers into slot-machine behavior: pull the lever, hope, repeat. Branching keeps every candidate alive, so choosing a direction is an editorial decision instead of a gamble. Q: Does branching work on existing novels, not just AI-written stories? A: Yes — that is arguably its best use. Import a finished or abandoned book, fork from any chapter, and explore the ending the author never wrote. The original text stays untouched; each what-if lives on its own line. Q: Doesn't keeping every branch get messy? A: Only if the tool hides structure. With a visible tree — where each line split, how far it has grown, which branch you are on — a dozen live branches stay navigable. The mess in single-line tools comes precisely from not having that map. ======================================================================== ## DOCS ### Foreverse Docs — App Guides and Creator Specs URL: https://foreverse.app/docs Official docs: module-by-module app guides (reader, roleplay, agent, models, settings) plus creator specs for chat beautification, chara_card_v3 import, worldbook runtime, and theme sideloading. ### Getting Started with Foreverse — First 10 Minutes URL: https://foreverse.app/docs/getting-started Set up a provider key (or the official channel), import your first book, run your first continuation, and start a character chat. ### Worlds and Shelf Guide — Import and Manage URL: https://foreverse.app/docs/guide-worlds Import txt / epub / character cards / worldbooks / .charx / group JSON, manage the shelf, and use the world detail tabs. ### Reader Guide — Selection AI, Options, Branches, Listening URL: https://foreverse.app/docs/guide-reader Every reader feature: 14 selection actions, the four continuation modes, story options, AI-segment editing, branch timeline, TTS listening, page-turn and theme settings. ### Roleplay Guide — Tavern Actions, Groups, IM Mode URL: https://foreverse.app/docs/guide-roleplay The full map for tavern players: 12 plus-panel actions, message-level tools, style presets, QR and automations, group strategies, IM mode, media. ### Agent and Companion Guide — Tasks, @, Skills URL: https://foreverse.app/docs/guide-agent Assign tasks, read the reasoning stream and tool cards, use @ references across books and cards, customize skills, and run a Companion. ### Models and Usage Guide — BYOK Setup and Request Records URL: https://foreverse.app/docs/guide-models Add providers, test connectivity, set per-modality defaults, understand official-channel billing, and read the API request record. ### Settings and Data Guide — Appearance, Assets, Export URL: https://foreverse.app/docs/guide-settings Themes and sideloading, the SillyTavern asset library, tavern preferences, data export and account deletion. ### Beautify Spec — Chat Cards, Stickers, HTML Rendering URL: https://foreverse.app/docs/beautify Creator reference: ```html WebView rendering and sandbox limits, fv-card native cards, [[sticker:]] syntax, the marker-regex-HTML pipeline, TavernHelper API support. ### Character Card Import Spec — PNG / JSON / .charx URL: https://foreverse.app/docs/character-cards Supported card containers and fields: PNG ccv3 chunks, JSON, V3 .charx zip with assets, V1-V3 compatibility, group JSON. ### Worldbook Spec — Triggering, Budget, V3 Decorators URL: https://foreverse.app/docs/worldbook Exact runtime behavior: keyword scanning and depth, constant entries, recursion, token-budget drops, and @@activate / @@dont_activate semantics. ### Glossary — Tavern, Character Cards, and AI Reading, Defined URL: https://foreverse.app/docs/glossary High-frequency terms defined in one place: character card / chara_card_v3 / lorebook / constant entries / recursion / scan depth / regex / presets / macros / swipe / BYOK / prompt caching / context window / branch / VN theater and more — 27 self-contained, quotable entries with further reading. ### Theme Sideload Spec — .fvtheme.json Fields URL: https://foreverse.app/docs/themes The .fvtheme.json format: top-level fields, all FvPalette semantic color keys, ARGB notation, fallback rules and design guidelines. ### Foreverse 创作者文档 — 美化规范 · 角色卡 · 世界书 · 主题 URL: https://foreverse.app/zh/docs 面向角色卡创作者的官方文档：聊天美化规范（fv-card / 贴纸 / HTML 渲染 / marker 正则链路）、chara_card_v3 导入规范、世界书 runtime 行为、.fvtheme.json 主题侧载格式。 ### Foreverse 快速上手 — 10 分钟从装好到第一次续写 URL: https://foreverse.app/zh/docs/getting-started 新手教程：配置第一个 API Key（或用官方渠道）、导入第一本 txt/epub、完成第一次划词续写、导入第一张角色卡并开聊。每一步都有位置指引。 ### 世界与书架教程 — 导入、管理与世界详情 URL: https://foreverse.app/zh/docs/guide-worlds 书架（世界 tab）完整教程：txt / epub / 角色卡 / 世界书 / .charx / 群聊 JSON 的导入方法，筛选与管理，世界详情的分支、角色、世界书三个页签怎么用。 ### 阅读器完整教程 — 划词 AI、续写表单、选项生成、分支与听书 URL: https://foreverse.app/zh/docs/guide-reader 阅读器全部功能逐项说明：14 项划词操作、续写表单四模式、剧情候选（选项生成）、AI 段编辑与继续、半自动模式、分支时间线、听书缓存与定时、五种翻页与主题设置。 ### 角色聊天完整教程 — 酒馆动作、世界书、群聊与线上聊 URL: https://foreverse.app/zh/docs/guide-roleplay 酒馆玩家的完整功能地图：+ 面板 12 动作、消息级操作（swipe/编辑/隐藏/checkpoint/分支）、写作风格预设、Persona、Quick Reply 与自动化、群聊策略、线上聊 IM 模式、图片/朗读/翻译。 ### 智能体与伴侣教程 — 任务、@ 引用、技能与陪伴 URL: https://foreverse.app/zh/docs/guide-agent 智能体 tab 使用教程：给 Agent 派任务、看懂思考流与工具卡、@ 引用小说和角色卡、自定义技能目录；以及伴侣（Companion）的记忆、贴纸、主动消息和搬家包。 ### 模型与用量教程 — BYOK 配置、按模态默认与请求记录 URL: https://foreverse.app/zh/docs/guide-models 从零配好模型：添加供应商、填 Key、测连通、启停模型、按文/图/音/视频设默认；官方渠道计费说明；API 请求记录里 tokens、缓存命中和费用怎么看。 ### 设置与数据教程 — 外观、酒馆资产库与数据导出 URL: https://foreverse.app/zh/docs/guide-settings 设置 tab 全项说明：主题外观与侧载主题、酒馆资产库（角色/世界书/预设/正则/QR/Persona）、酒馆偏好、数据导出与账号删除。 ### Foreverse 美化规范 — 聊天卡片、贴纸与 HTML 渲染 URL: https://foreverse.app/zh/docs/beautify 创作者美化手册：```html 围栏 WebView 渲染与沙箱边界、```fv-card 原生小卡三型、[[sticker:]] 贴纸语法、marker → 正则 → 主题化 HTML 的整卡美化链路、TavernHelper 宿主 API 支持清单。 ### Foreverse 角色卡导入规范 — PNG / JSON / .charx URL: https://foreverse.app/zh/docs/character-cards 支持的角色卡格式与字段：PNG ccv3/chara 隐写、JSON、V3 .charx zip 容器（含资产分类）、V1/V2/V3 字段兼容、nickname、群聊 JSON 导入。 ### Foreverse 世界书规范 — 触发、预算与 V3 decorators URL: https://foreverse.app/zh/docs/worldbook 世界书 runtime 行为参考：关键词触发与扫描深度、常驻条目、递归激活、token 预算丢弃规则，以及 V3 decorators（@@activate / @@dont_activate）的确切语义。 ### 术语表 — 酒馆、角色卡与 AI 阅读的黑话手册 URL: https://foreverse.app/zh/docs/glossary 高频术语集中定义：角色卡/chara_card_v3/世界书/常驻条目/递归/扫描深度/正则/预设/宏/swipe/BYOK/prompt 缓存/上下文窗口/分支/剧场模式等 27 条，每条自包含可引用，附站内延伸阅读。 ### Foreverse 主题侧载格式 — .fvtheme.json 完整字段 URL: https://foreverse.app/zh/docs/themes .fvtheme.json 主题包格式：顶层字段、FvPalette 全部语义色键、ARGB 颜色写法、回退规则与设计准则；文件放入 /foreverse/themes/ 即在设置 → 外观出现。 ======================================================================== Contact: support@foreverse.app Get the app: https://play.google.com/store/apps/details?id=app.foreverse.android