AI UP LAB面向任务的 AI 实践资源AI resources for real tasks
收藏Saved
更多More
StarStar
HAL-X AI HAL-X AI 官方 Official 社区 showcase Community showcase

THX-01：322M 开源决策模型，一次前向给出带校准概率的选项 THX-01: a 322M open decision model that returns calibrated choices in one forward pass

THX-01: a 322M open decision model that returns calibrated choices in one forward pass

HAL-X AI（阿塞拜疆）在 2026-10-10 开源的 THX-01：约 3.22 亿参数的非自回归决策模型，不写长文，只对你给出的现状和若干类型化问题（选择题、是否、打分、抽数字、摘原文、找依据）一次前向返回带校准概率的结构化答案。权重与 Python 包 `thx01` 为 Apache-2.0；另有托管 API 与 Hugging Face Space 演示。作者在 HF 博文里用内部多语工单基准对比 TypeSafe Jev，并给出延迟等数字——本库没有复跑。适合路由、分流、抽取这类「只要一个决定」的 Agent 环节。 THX-01 from HAL-X AI (Azerbaijan), open-sourced 2026-10-10: a ~322M non-autoregressive decision model that does not write long prose. You pass a state plus typed questions (choice, yes/no, score, number extract, excerpt, citation) and get structured answers with calibrated probabilities in one forward pass. Weights and the Python package `thx01` are Apache-2.0; there is also a hosted API and a Hugging Face Space demo. The HF blog compares it to TypeSafe Jev on an internal multilingual ticket benchmark with latency figures—this library did not re-run them. Good for agent steps that only need a decision: routing, triage, extraction.

预览截自 Hugging Face 博文页 Preview cropped from the Hugging Face blog page

来源Sources

官方模型页 Official model page Hugging Face 博文：THX-01 开源说明 Hugging Face blog: THX-01 open-source announcement Farid Aghayev (doofz) 2026年10月10日 Oct 10, 2026 https://huggingface.co/blog/doofz/thx-01-we-open-sourced-a-322m-decision-model-that 官方模型页 Official model page Hugging Face 模型：doofz/THX-01（Apache-2.0） Hugging Face model: doofz/THX-01 (Apache-2.0) doofz / HAL-X AI 2026年10月10日 Oct 10, 2026 https://huggingface.co/doofz/THX-01 官方模型页 Official model page 托管 API 文档 Hosted API docs HAL-X 2026年10月10日 Oct 10, 2026 https://api.hal-x.ai/docs/thx-01/ 产品主页 Product home Hugging Face Space 演示 Hugging Face Space demo doofz 2026年10月10日 Oct 10, 2026 https://huggingface.co/spaces/doofz/THX-01-demo 社区 showcase Community showcase Show HN / X 线索（开源决策模型） Show HN / X pointer (open decision model) lebagetdefrance 2026年10月11日 Oct 11, 2026 https://news.ycombinator.com/item?id=50048015

为什么推荐Why we recommend it 与 Jev 的对比数字来自作者内部基准，本库没有安装 thx01，也没有复跑。收录进 Models，与库内 Strands Decider、Jeff、Clef、Jev 同属「决策小模型」一条线。 Jev comparison numbers are the authors’ internal benchmarks; this library did not install thx01 or re-run them. Listed in Models alongside Strands Decider, Jeff, Clef, and Jev as another small decision model.

核验记录：来源、抓取与截图说明Verification log: sources, checks, and captures

主源是 2026-10-10 HF 博文与模型卡 doofz/THX-01，早间 HTTP 200；能力表、Apache-2.0、pip 与 API 链接照博文。预览为 Playwright 实拍博文页裁 3:2 WebP。Primary sources are the 2026-10-10 HF blog and model card doofz/THX-01, HTTP 200 this morning; capability table, Apache-2.0, pip, and API links follow the blog. Preview is a Playwright capture of the blog page, cropped to 3:2 WebP.

要点（官方称 / 社区称）Highlights (vendor / community claims)

同板相关Related on this board

同一看板上互相关联的卡：社区 demo 指回它跑的模型，模型指向值得看的 demo。Cards linked on this board: a community demo points back to the model it ran on; the model points to demos worth a look.

Strands Agents Strands Agents 官方 Official 社区 showcase Community showcase

Strands Decider 2B：一个小的开源决策模型，只做选择题和打分 Strands Decider 2B: a small open decision model that picks an option or gives a score

Strands Decider 2B: a small open decision model that picks an option or gives a score

Strands Decider 2B 是 Strands Agents 放出的一个小决策模型。它不写文章。你给它一段现状和几个允许的答案，它为每个答案打一个概率，也可以按一个分数尺打分。仓库写它大约 19 亿参数，基于 Qwen3.5-2B 改来，拿掉了原来用来生成文字的那一层，换上一个大约一百万参数的小头。代码和权重以 Apache-2.0 发布，可以在自己的 CPU 或显卡上跑。官博是 2026-10-01，仓库现在指向的参考版本是 v21，更早的 v19 仍在。本库没有安装，也没有复跑仓库里的分数。 Strands Decider 2B is a small decision model from Strands Agents. It does not write prose. You give it a situation and the answers you allow, and it returns a probability for each one, or a score on a scale. The repo says it has about 1.9 billion parameters, starts from Qwen3.5-2B, removes the layer that used to generate text, and adds a small head of about a million parameters. Code and weights are Apache-2.0, and it can run on your own CPU or GPU. The blog is dated 2026-10-01. The repo’s reference version now is v21, and the earlier v19 is still published. This library did not install it and did not re-run the repo’s scores.

厂商公告Vendor announcement 2026年10月1日 Oct 1, 2026 4 官方 · 1 社区 4 official · 1 community
kouhxp kouhxp 官方 Official 社区 showcase Community showcase

Gutsy：本机 CPU 可跑的 0.8B Jev 兼容决策模型 Gutsy: 0.8B Jev-compatible decision model for local CPU

Gutsy: 0.8B Jev-compatible decision model for local CPU

kouhxp 开源的 gutsy-0.8b（v0.4）：把 Qwen3.5-0.8B 微调成决策模型，不生成自由文本。你送一段状态（文本或 JSON）和 typed questions（yes/no、choice、score），它为每个选项返回校准概率；choice / score 还带 reject。运行时 gutsy-inference 走 llama.cpp，默认可在普通 CPU 上起本地 HTTP；接口刻意对齐 Jev / System One（`/v1/systemone` 等），现有客户端多半只换 base URL。权重在 Hugging Face kouhxp/gutsy（Q8_0 与 Q4_K_M GGUF），代码与权重 README 写 Apache-2.0。JevBench / 评测表与延迟是作者自测，本库未复跑。同板与 Jev、Laya、Clef、Jeeves、Ollaya 相邻；Cases 已有本机 Jeff 与 Onesie CLI。 kouhxp’s open gutsy-0.8b (v0.4) fine-tunes Qwen3.5-0.8B into a decision model that does not generate free text. You send a state (text or JSON) plus typed questions (yes/no, choice, score) and get a calibrated probability for every option; choice / score answers also carry reject. The gutsy-inference runtime uses llama.cpp and can serve a local HTTP API on an ordinary CPU. The API deliberately matches Jev / System One (`/v1/systemone` and related), so many existing clients only need a new base URL. Weights live on Hugging Face kouhxp/gutsy (Q8_0 and Q4_K_M GGUF); README says Apache-2.0 for code and weights. JevBench / eval tables and latency are author self-tests, not re-run here. On this board it sits next to Jev, Laya, Clef, Jeeves, and Ollaya; Cases already lists local Jeff and the Onesie CLI.

厂商公告Vendor announcement 2026年10月1日 Oct 1, 2026 2 官方 · 1 社区 2 official · 1 community
Cloudflare Cloudflare 官方 Official 社区 showcase Community showcase

Clef：Cloudflare 开源决策模型（含 Clef-Flash） Clef: Cloudflare open-weight decision models (incl. Clef-Flash)

Clef: Cloudflare open-weight decision models (incl. Clef-Flash)

Cloudflare 发布开源决策模型 Clef，以及更小的 Clef-Flash。它们不生成自由文本：读入一段状态和一组带类型的问题，为每个允许的答案返回概率。Workers AI 上的模型 id 是 @cf/cloudflare/clef（27B）和 @cf/cloudflare/clef-flash（9B）。官方称接口与 Jev / System One 同形，问题类型仍是 noul / choice / score，并带视觉编码器，状态可以是文本、JSON、图像或视频。权重在 Hugging Face，许可 Apache-2.0。HF 卡写 Clef 由 Qwen/Qwen3.8-27B 后训练、Clef-Flash 由 Qwen/Qwen3.5-9B 后训练，这是厂商 / HF 口径。官博还宣布 RL 微调：先由 forward-deployed engineer 动手，之后再做成自助平台——按路线图记，不是已经核过的产品深度。Decision Index、延迟和工作流分数一律是厂商自述，本库未复跑。同板与 Jev、Laya、Jeeves、Ollaya 相邻；Cases 里已有本机 Jeff 和命令行 Onesie。 Cloudflare released the open-weight decision model Clef and the smaller Clef-Flash. They do not generate free text: they read a state plus typed questions and return a probability for every allowed answer. On Workers AI the ids are @cf/cloudflare/clef (27B) and @cf/cloudflare/clef-flash (9B). The vendor says the API matches Jev / System One, with noul / choice / score questions and a vision encoder, so state can be text, JSON, images, or video. Weights are on Hugging Face under Apache-2.0. The HF cards say Clef is post-trained from Qwen/Qwen3.8-27B and Clef-Flash from Qwen/Qwen3.5-9B — a vendor / HF claim. The blog also announces RL fine-tuning: hands-on with a forward-deployed engineer first, then a later self-serve platform. That is a roadmap, not a product depth this library verified. Decision Index, latency, and workflow scores are vendor claims and were not re-run here. On this board it sits next to Jev, Laya, Jeeves, and Ollaya; Cases already lists local Jeff and the Onesie CLI.

厂商公告Vendor announcement 2026年10月1日 Oct 1, 2026 4 官方 · 1 社区 4 official · 1 community
TypeSafe AI TypeSafe AI 官方 Official 社区 showcase Community showcase + 精选 + Featured

Jev / System One Models（TypeSafe AI） Jev / System One Models (TypeSafe AI)

Jev / System One Models (TypeSafe AI)

TypeSafe AI 于 2026-09-15 发布的首个公开 System One Model。面向软件内决策而非聊天：输入非结构化状态，输出类型安全的结构化决策，并附带校准概率与置信度。官方称新架构 + 并行采样 + Reinforcement Learning for Calibrated Decisions（RLCD）；目前为早期访问；速度、价格与评测倍率均为厂商自述，本库未复跑其工作流评测。 TypeSafe AI’s first public System One Model, announced 2026-09-15. Built for in-software decisions rather than chat: unstructured state in, type-safe structured decisions out with calibrated probabilities and confidence. Vendor claims a new architecture, parallel sampling, and Reinforcement Learning for Calibrated Decisions (RLCD); currently early access; speed, pricing, and eval multipliers are all vendor claims, and the workflow evals were not re-run by this library.

厂商公告Vendor announcement 2026年9月15日 Sep 15, 2026 2 官方 · 1 社区 2 official · 1 community
PostHog PostHog 官方 Official 社区 showcase Community showcase + 精选 + Featured

Jeeves：带推理的开源 Jev 风格决策模型（PostHog） Jeeves: open Jev-style decision model with reasoning (PostHog)

Jeeves: open Jev-style decision model with reasoning (PostHog)

HN https://github.com/PostHog/jeeves — MIT：PostHog 开源的「会先想再决定」的 Jev 风格分类器。README：基于 Qwen3.5-9B（LoRA + pointer head），SFT + CISPO 训练，带 block-4 diffusion drafter；同一请求支持 noul / choice / score，API 与 Jev 同形（`/v1/systemone`）。作者表（thinking、greedy、2560-token 上限；Kev-9B / Jev 列为 Kev 公布数字）：Test overall 0.889 vs Kev 0.822 / Jev 0.857；JevBench overall（231 public）0.935 vs Jev 0.866；JevBench hard 0.865 vs Jev 0.730。延迟自述：无 thinking 约 0.3 s/请求，有 thinking 中位约 3.3 s（单 H100）；需 CUDA（Hopper 才用 FP8 kernel）。权重 https://huggingface.co/PostHog/jeeves。README 写 Inspired by Kev。公开仓与 HF 晚间 GET 200。本库未复跑基准、未代下权重。 README: Qwen3.5-9B with LoRA + pointer head, trained with SFT + CISPO, plus a block-4 diffusion drafter; same request supports noul / choice / score via a Jev-shaped /v1/systemone API. Author table (thinking, greedy, 2560-token cap; Kev-9B / Jev columns are Kev-published): Test overall 0.889 vs Kev 0.822 / Jev 0.857; JevBench overall (231 public) 0.935 vs Jev 0.866; JevBench hard 0.865 vs Jev 0.730. Latency claim: about 0.3 s/request without thinking, ~3.3 s median with it on one H100; needs CUDA (Hopper for the FP8 kernel). Weights at https://huggingface.co/PostHog/jeeves. README says Inspired by Kev. Public repo and HF evening GET 200. Benchmarks not re-run here; weights not downloaded for you.

厂商公告Vendor announcement 2026年9月29日 Sep 29, 2026 2 官方 · 1 社区 2 official · 1 community
ConvAI Innovations ConvAI Innovations 官方 Official 社区 showcase Community showcase

Laya：开源 System 1 决策引擎（对照 Jev） Laya: open-source System 1 decision engine (Jev-adjacent)

Laya: open-source System 1 decision engine (Jev-adjacent)

作者称 Laya 是非自回归、亚 35ms 的开源 System 1 决策模型族（Apache 2.0 权重），提供 choice / score / noul 三类结构化决策与多语路由；对照站内已收录的 TypeSafe Jev。基准、延迟、与 Jev 对比数字一律标「作者称 / 第三方整理」，本库未复跑。HN 讨论：https://news.ycombinator.com/item?id=49765348（2026-09-19）。 Author presents Laya as an open-weight non-autoregressive System 1 decision family (Apache 2.0) with choice/score/noul primitives and multilingual routing; sits next to catalogued TypeSafe Jev. Latency/benchmark/Jev-comparison figures are author or third-party claims — not re-run here.

厂商公告Vendor announcement 2026年9月19日 Sep 19, 2026 1 官方 · 2 社区 1 official · 2 community