Skip to content

Lynn v0.87.1

Latest
Latest

Choose a tag to compare

View all tags
@MerkyorLynn MerkyorLynn released this 11 Oct 14:48
v0.87.1
d37035e

国内镜像站下载（推荐） / Downloads

国内用户请优先使用镜像站；GitHub Assets 作为备用下载。

npm install -g --force "https://download.merkyorlynn.com/downloads/cli/lynn-cli-0.87.1.tgz"
Lynn --version

Lynn v0.87.1 Release Notes / 发布说明

2026-10-11 · Step 5、实时语音与 Coder390 本地模型

中文重点

  • 安装容量门槛：低于 14GiB（如 13GB）或检测失败时禁止安装；14～不足 16GiB 仅达到 Q2 最低门槛，16GB 推荐 Q2，24GB 及以上推荐 Q3。Apple Silicon 看统一内存，NVIDIA 看单卡显存，不叠加多卡、不混用系统内存。手动选档不能绕过检查，文件之外还需预留上下文缓存和运行缓冲。

  • GUI / CLI 默认工作模型：统一使用 step-5-preview，保留已有配置的兼容入口。

  • 原生视觉与复查：视觉验收跟随当前默认模型，不再固定 MiMo；修复云端原生图片保留和手动复查遗漏设备鉴权的问题。真实识图、手动复查和并发复查均纳入安装包门禁，失败降级不算通过。

  • 实时语音：恢复 stepaudio-2.5-realtime，GUI 与 CLI 共用 Brain 托管的语音链路。

  • 工具参数由模型分析：天气、体育、行情和汇率接收模型填写的地点、联赛、日期、球队、代码和货币对；取消从用户整句问题猜参数的预取路径。体育数据源不再猜日期或从句子抽球队，缺日期时明确要求模型补齐。

  • 复合地点：保留“顺德大良”等完整地点；必要的上级地点由模型明确提供，不自动裁剪或猜测。

  • 模型回答收尾：不再把“本次证据不足或上游失败”误判为没有工具能力。候选耗尽时只做一次有界模型收尾；原始工具账本不再冒充成功答案。

  • 客户端收尾修复：工具轮数用尽后让原模型基于完整证据做无工具收尾，不再裁剪搜索片段拼答案。持续思考或输出会刷新工具后空闲计时，总时限保持不变。

  • 纯整理任务：GUI 与 CLI 共用任务遵循提示，保留用户原文的相对日期与全部行动项，不擅自增加日历查询；需要查证、换算或实时信息时仍由模型决定工具参数。

  • Coder390 本地模型：Qwen3.8-27B Coder390 EfficientThink Q3 / Q2，来自指定 ModelScope GGUF 仓库。Q3 为 17.30 GB、8K 初始上下文；Q2 为 13.28 GB、4K 初始上下文。24GB 及以上默认选择 Q3，24GB 以下（16GB 档）选择 Q2；实际运行还需缓存与缓冲空间。

  • 内置 Q4 MTP：两档使用内置 MTP，不另下载 DFlash2 草稿；本次安装不含视觉组件。按固定字节数与 SHA-256 校验文件，保留已经下载的旧模型和手动导入入口。

  • 设置与说明对齐：默认安装选项保留 Q3 / Q2，更新五种语言提示、中英文 README 和运行时帮助。

  • 固定数据输出：要求输出指定 JSON 或字面数据时，运行时信息不得改写用户给定的常量值。

  • 延续 V0.87.0 的窗边树影、斜杠模板、文件与图片面板，以及按需安装的专家圆桌插件。

English highlights

  • Block installation below 14GiB or when hardware detection fails. Q2 is recommended at 16GB, Q3 at 24GB; 14 to under 16GiB only meets Q2's minimum. Detect Apple Silicon unified memory or the largest single NVIDIA GPU's VRAM, not summed GPUs or host RAM. Manual selection cannot bypass memory checks.

  • GUI and CLI request step-5-preview by default, retaining migration compatibility.

  • Vision acceptance follows the selected default model rather than requiring MiMo. Preserve native image inputs in Brain and attach device authentication to manual review sessions. Real image recognition, manual review and concurrent review are required package checks; failure fallbacks do not count as passes.

  • Restore stepaudio-2.5-realtime for the shared Brain-hosted realtime voice path.

  • Let the model supply places, leagues, dates, teams, tickers and currency pairs. Remove entity guessing from whole-question prefetch; sports sources require explicit dates instead of guessing dates or teams from sentences.

  • Preserve compound places such as Shunde Daliang; parent-place fallbacks must be supplied explicitly by the model.

  • Preserve model-written explanations of missing evidence or upstream failures instead of misclassifying them as missing tool capability. Exhausted routes use one bounded model recovery; raw tool ledgers no longer masquerade as successful answers.

  • Let the original model finalize complete tool evidence without further tools, rather than constructing answers from clipped snippets. Active reasoning or text refreshes the post-tool idle timer without extending the hard turn deadline.

  • Share task-fidelity guidance across GUI and CLI: preserve supplied relative dates and every action item during pure transformations, without unrequested calendar lookups. Requested verification, date conversion and realtime research still use model-chosen tool parameters.

  • Switch local tiers to Qwen3.8-27B Coder390 EfficientThink: Q3 is 17.30 GB with an initial 8K context; Q2 is 13.28 GB with an initial 4K context. The default tier selection is Q3 at 24GB and above, Q2 below 24GB (16GB tier). Caches and runtime buffers require additional memory.

  • Both GGUF tiers include Q4 MTP, with no separate DFlash2 draft or vision component downloaded by this installation. Verify pinned sizes and SHA-256 values; preserve older files and manual import.

  • Align settings, five-language hints, runtime help and both READMEs.

  • Preserve user-supplied constants when emitting requested JSON or other literal data.

  • Retain V0.87.0 interface features and the optional expert-roundtable plugin.

模型与仓库 / Models and repositories

Assets 6
Loading