社区 Community 需审批 Needs approval
Jeff — 本机可跑的 Jev 兼容 System One 小决策模型 Jeff — local Jev-compatible System One decision models
Jeff — local Jev-compatible System One decision models
在本机试用面向小决策的模型。 Try a small-decision model locally.
详细介绍Details
HN 首页 https://github.com/firelex/jeff — MIT：把 Qwen3.5 / Gemma 4 微调成零样本分类决策模型，请求格式与 TypeSafe Jev 同形；作者称约 22 ms（RTX PRO 6000）/ 28 ms（Apple M4 Max MLX）一次前向，输出是各选项校准概率，不是生成文本。README：`jeff-serve` + `/v1/systemone`，支持 choice / noul / score；训练全在本机硬件，合成数据来自开源模型。Hugging Face 有 Jeff-Qwen3.5-0.8B / 2B 与 Jeff-Gemma4-E2B。作者写明独立项目、非 TypeSafe 官方。本条是工具笔记，不替你下载权重或代跑服务。 The author reports about 22 ms (RTX PRO 6000) / 28 ms (Apple M4 Max MLX) per forward pass, returning calibrated option probabilities rather than generated text. README: jeff-serve plus /v1/systemone with choice / noul / score; training on local hardware; synthetic data from an open model. Hugging Face hosts Jeff-Qwen3.5-0.8B / 2B and Jeff-Gemma4-E2B. The author states it is independent and not affiliated with TypeSafe. This is a tool tip — it does not download weights or run the server for you.
它是什么What it is
Jeff 是 MIT 开源的本机 System One 决策模型线（仓库 firelex/jeff）。作者把 Qwen3.5 与 Gemma 4 微调成零样本分类器：你用自然语言描述状态和选项，模型一次前向返回各选项校准概率（choice / noul / score），请求格式刻意对齐 TypeSafe Jev。README 给 `jeff-serve` 与 `/v1/systemone` 的 curl 示例，并自称约 22–28 ms 级延迟（视硬件而定）。训练与合成数据流程写明跑在本机硬件、合成数据来自开源模型。Hugging Face 有 0.8B / 2B / Gemma4-E2B 权重。Show HN/HN 首页 item 49883844。本站收成「工具笔记」：给已在玩 Jev / Onesie 的人一条可本机跑的相邻栈，不是 Grok Bot 本体，也不是 TypeSafe 官方。 Jeff is an MIT-licensed local System One decision-model line (repo firelex/jeff). The author fine-tunes Qwen3.5 and Gemma 4 into zero-shot classifiers: describe a state and options in plain words, get calibrated option probabilities from one forward pass (choice / noul / score), in a request shape deliberately aligned with TypeSafe Jev. The README shows jeff-serve and a /v1/systemone curl example, and claims roughly 22–28 ms latency depending on hardware. Training and synthetic-data steps are described as local-hardware / open-model. Hugging Face hosts 0.8B / 2B / Gemma4-E2B weights. HN item 49883844. Listed here as a tool tip next to catalogued Jev / Onesie — not Grok Bot itself, and not TypeSafe-official.
为什么推荐Why we recommend it 收录理由：与本站已收的 Jev / Onesie 同属 System One 决策线，但可本机跑、可脚本化 HTTP。边界：基准表与毫秒数是作者 README / 自测口径，本站未复跑；作者自述在推理向基准上仍明显弱于 Jev；独立非 TypeSafe 官方。下载权重与起服务会占磁盘与 GPU/CPU，本站未代装。 Listing rationale: same System One decision line as catalogued Jev / Onesie, but runnable locally over a scriptable HTTP API. Boundary: benchmark tables and millisecond figures are the author’s README / self-test numbers, not re-run here; the author says reasoning-heavy benches stay well below Jev; independent, not TypeSafe-official. Downloading weights and serving uses disk and GPU/CPU; this site did not install it.
核验记录：来源、抓取与截图说明Verification log: sources, checks, and captures
主源是 HN 49883844 与 MIT 仓库 / Hugging Face 权重页。预览 Playwright 实拍 GitHub 仓库首屏裁 3:2。不打星。 （item 49883844，2026-09-28T20:23:36Z / 约 2026-09-29 04:23 CST，Algolia 核对 235pts） 仓库 firelex/jeff 于 2026-09-29 gh api：MIT、282★；Primary sources are HN 49883844 and the MIT repo / Hugging Face weight pages. Preview is a Playwright 3:2 capture of the GitHub repo above the fold. Unrated. (2026-09-28T20:23:36Z / about 2026-09-29 04:23 CST, 235pts on Algolia at check time) HN front page item 49883844 at https://github.com/firelex/jeff — MIT fine-tunes of Qwen3.5 / Gemma 4 for zero-shot classification in the same request shape as TypeSafe Jev. Repo firelex/jeff on the 2026-09-29 gh api: MIT, 282★;
对照Contrast
左列是 README / HN 对得上的事实。右列说明为什么放在 Cases 工具笔记，以及不要把它当成已经替你验证过的 Jev 替代品。 Left column: what the README / HN actually say. Right column: why it sits as a Cases tool tip, and why it is not a verified Jev replacement.
Jeff Jeff 对照本站口径 This library’s framing
是什么 What it is
Jeff Jeff
本机可服务的小决策模型：同一请求里多题，stdout/HTTP 拿概率与选项。 Small decision models you can serve locally: several questions in one request, probabilities and picks over HTTP/stdout.
对照本站口径 This library’s framing
Cases「AI tool tip」。相邻 Jev / Onesie；作者写明独立、非 TypeSafe 背书。 An AI tool tip in Cases. Adjacent to Jev / Onesie; the author states it is independent and not TypeSafe-endorsed.
速度与准度 Speed and accuracy
Jeff Jeff
README：毫秒级前向；总表接近或局部超过已发表的 Jev 数字，推理向基准仍明显更弱。 README: millisecond forward passes; overall tables approach or locally beat published Jev figures, while reasoning-heavy benches stay weaker.
对照本站口径 This library’s framing
数字是作者自测快照。本站未复跑；选型仍要你自己在任务上测。 Figures are the author’s self-test snapshot. This site did not re-run them; validate on your own task.
你要自己做的事 What you still do
Jeff Jeff
uv sync、下 Hugging Face 权重、起 jeff-serve，再 curl /v1/systemone。 uv sync, download Hugging Face weights, start jeff-serve, then curl /v1/systemone.
对照本站口径 This library’s framing
本站未代下权重、未代起服务。工单原文与密钥不要当默认 payload。 This site did not download weights or start a server. Do not send raw tickets or secrets as the default payload.
怎么试Try it
- 先读仓库和 HN Read the repo and the HN thread
打开 GitHub README 与 item 49883844，确认 MIT、Jev 兼容口径、以及「独立非 TypeSafe」声明。 Open the GitHub README and item 49883844. Confirm MIT, the Jev-compatible framing, and the independent / not-TypeSafe statement.
- 按作者方式起本地服务 Start the local server the author’s way
README：`uv sync`，`hf download mstrasser/Jeff-Qwen3.5-0.8B`，再 `JEFF_CHECKPOINT=… uv run jeff-serve`（Apple 可用 MLX extra）。先对本机 curl，再接到你的脚本。 README: uv sync, hf download mstrasser/Jeff-Qwen3.5-0.8B, then JEFF_CHECKPOINT=… uv run jeff-serve (MLX extra on Apple). Curl locally before wiring your scripts.
- 用真实任务抽检，别只看总表 Spot-check on your task — not only the overall table
作者也写：零样本不够就短微调；推理向基准仍弱于 Jev。把你的路由 / 审核样例跑一遍再决定是否进流水线。 The author also says a short fine-tune helps when zero-shot is not enough, and reasoning benches stay below Jev. Run your own routing / moderation samples before putting it in a pipeline.
