Lynn v0.85.7 Release Notes / 发布说明

发布日期: 2026-07-07 · 27B Coding Q4 MTP 默认端侧模型 / 本地模型安装链路修复 / 三远端+镜像发版纪律

国内镜像站下载（推荐）

国内用户请优先使用以下镜像站地址；正式版本记录见 GitHub / Gitee Releases，下载请以镜像站为准。

GitHub Releases (old): https://github.com/LynnMerkyor/Lynn/releases/tag/v0.85.7
GitHub Releases (new): https://github.com/MerkyorLynn/Lynn/releases/tag/v0.85.7
Gitee Releases: https://gitee.com/merkyor/Lynn/releases/tag/v0.85.7

CLI:

npm install -g --force "https://download.merkyorlynn.com/downloads/cli/lynn-cli-0.85.7.tgz"

macOS Apple Silicon / ARM64: https://download.merkyorlynn.com/downloads/Lynn-0.85.7-macOS-arm64.dmg
macOS Intel / x64: https://download.merkyorlynn.com/downloads/Lynn-0.85.7-macOS-x64.dmg
Windows x64: https://download.merkyorlynn.com/downloads/Lynn-0.85.7-Windows-Setup.exe
下载页: https://download.merkyorlynn.com/download.html

端侧默认模型换成 27B Coding Q4 MTP: 本地模型推荐改为公开可下载的 Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-GGUF Q4 imatrix MTP 四分片，默认下载 Q4_LynnStyle/Q4-imatrix-MTP-00001-of-00004.gguf 到 00004，总量约 19.6GB。
公开模型链接修正: README、下载页、Release notes 和模型入口不再指向打不开的非 GGUF 页面，ModelScope / HuggingFace 都指向 GGUF 镜像。
低配不主动打扰: 硬件不足时聊天窗不会主动弹 27B 安装引导；9B / 4B 只保留为设置页里的低配降级选择。
修复 GitHub #75 本地模型 python3_not_found: 桌面端本地模型安装优先走 Electron 主进程 downloader，不再把用户带回旧 Python bootstrap；缺 Python 不再阻断“设置 → 模型 → 加载本地模型”。
下载链路支持四分片校验: 27B Q4 默认档会按 shard 顺序下载、校验 SHA256、保留已完成分片，并在完成后按用户授权启动 llama.cpp MTP 端点。
发版流程加严: 每次新版本/同版本热修都必须同步 LynnMerkyor/Lynn、MerkyorLynn/Lynn、Gitee merkyor/Lynn 和腾讯镜像下载站；release:verify-remotes 默认校验三个代码远端。
保留 v0.85.6 修复: 本地绝对路径读取、file:// 元问题、防串题污染、Windows CMD 弹窗、会话进度右栏和 GUI/CLI 回归门禁继续保留。

npm test -- tests/llamacpp-profiles.test.js desktop/src/react/settings/tabs/providers/local-qwen-provider.test.ts tests/migrate-providers.test.js 通过。
npm test -- desktop/__tests__/ProviderDetail.helpers.test.ts 通过。
node --check 通过：desktop/llamacpp-profiles.cjs、desktop/local-model-controller.cjs、desktop/llamacpp-manager.cjs、desktop/model-downloader.cjs、scripts/verify-release-remotes.mjs。
gate:cli-100 通过：100/100。
macOS arm64 / x64 DMG 已完成应用公证、最终 DMG 公证、staple 和 Gatekeeper 校验。
Windows x64 NSIS 安装包已生成并签名；packaged server / packaged CLI smoke 通过。
镜像站 0.85.7 资产公网 URL 已验证，远端 CLI tarball sha256 与本地一致，test:cli-install:remote 通过。
release:verify-remotes 通过：LynnMerkyor/Lynn、MerkyorLynn/Lynn、Gitee merkyor/Lynn 的 main 和 v0.85.7 已同步；两个 GitHub 仓库 CI 均通过。

Release date: 2026-07-07 · 27B Coding Q4 MTP default edge model / local-model setup fix / three-remote + mirror release discipline

Switches the default local edge model to 27B Coding Q4 MTP: Lynn now recommends the public Qwen3.6-27B-DSV4Pro-GLM52-SFT-GPT55-RL-Coding-GGUF Q4 imatrix MTP split GGUF by default, about 19.6GB across four shards.
Fixes public model links: README, release notes, download pages, and model entry points now link to the public GGUF mirrors instead of unavailable non-GGUF pages.
Keeps low-config machines quiet: Lynn no longer proactively prompts underpowered devices to install 27B; 9B / 4B remain explicit downgrade choices in Settings.
Fixes GitHub #75 python3_not_found: desktop local-model setup now uses the Electron main-process downloader first, so missing Python no longer blocks Settings → Models → load local model.
Supports four-shard download and verification: the 27B Q4 default downloads each shard, verifies SHA256, preserves completed shards, and starts the llama.cpp MTP endpoint only after user authorization.
Hardens release discipline: every release or same-version hotfix must update both GitHub repos, Gitee, and the Tencent download mirror; release:verify-remotes checks all three code remotes by default.