[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-amazon-strands-decider-2b":3,"topics-all":38,"news-related-ca3472cb-0638-440d-ae0e-f6c3eac95cf7":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"ca3472cb-0638-440d-ae0e-f6c3eac95cf7","亚马逊开源2B决策模型:权重、数据、训练脚本全放","亚马逊 AWS 旗下 Strands 团队开源 2B 决策模型:砍掉 Qwen3.5-2B 的语言建模头换成指针头,只做选项判断与打分,单次前向、中位延迟 115 毫秒。权重、训练数据与脚本全部放出,单张 RTX 3090 约 11 小时可重训。决策模型赛道两周内已有四家厂商入场。","亚马逊的 agent 团队 strands-labs 在 10 月 1 日开源了一个不太一样的模型:Strands Decider 2B。它不写代码、不聊天、不做摘要——只负责在给定选项里挑一个,或者打个分。权重、训练数据、训练脚本全部放上了 Hugging Face 和 GitHub,Apache 2.0 协议,本地 CPU 或 GPU 都能跑。\n\n## 先说品类:决策模型是什么\n\n这波热潮的起点是 TypeSafe AI 本月早些时候发布的 Jev。决策模型(decision model,也叫 \"system one\" 模型)和 LLM 的分工很清晰:LLM 生成任意文本,决策模型只做三类事——noul(是非题)、choice(N 选一)、score(按等级打分)。作为交换,它在同等参数量下更快,永远从给定选项里出答案,而且每个决策都带一个校准过的置信度——官方特意强调,这是各家前沿 LLM 推理 API 都不提供的东西。\n\n官方博客给了个直白的例子:一句「Help! My payouts have been failing for 3 days!」,问「该哪个团队处理?billing\u002Fsales\u002Fretail」,模型答 billing、置信度 0.768,并把三个选项的概率全部列出。\n\n## 架构:砍掉语言头,装一个指针头\n\nStrands Decider 2B 的做法很外科手术:拿 Qwen3.5-2B 预训练躯干,把语言建模头整个删掉——模型从此失去生成文本的能力——换上一个约 100 万参数的 pointer head。这个头拿 `\u003Canswer>` 位置的隐状态去和每个选项位置的隐状态对比打分,一次前向传播出结果,没有解码循环。躯干只用 rank-16 的 LoRA 微调。\n\n因为这个头不持有任何按选项区分的参数,模型学不到「第一个选项通常对」这种捷径,选项数量也不受架构限制——标签集合由请求定义,不烧进权重。三种问题类型(noul\u002Fchoice\u002Fscore)都从同一个 masked softmax 里读出来。\n\n数字方面:JevBench 公开集 231 个任务,accuracy 0.723(167\u002F231);分层看 easy 1.000、standard 0.875、hard 0.505;置信度 0.9 以上的答案约 95% 正确。延迟上,RTX 3090 中位 115 毫秒、95 分位 299 毫秒;M3 MacBook 跑 300 token 以内的小任务中位 153 毫秒。按官方引用的第三方榜单,它在 2B 级 33 个模型里排第 3,排除略超 2B 的模型后在 30 个里排第 1。\n\n还有个彩蛋:Qwen3.5-2B-Base 本身原生多模态,服务端加 `--vision` 保留视觉塔后,同一个 v19 checkpoint 不重训就能对图像做判断,官方测得 NaturalBench 上校准误差 ECE 0.014,比专门拿图像训练的 2B decider(0.080)还好。\n\n## 全配方开源,连失败实验都记录\n\n这个仓库最特别的不是模型,是过程。训练成本压到「人人可复现」的量级:单张 RTX 3090 约 11 小时重训整个 recipe,8 张 H100 开 FAST 模式 1 小时 10 分钟跑完。数据源许可逐条列在 data\u002Fsources.md,复现契约写在 data\u002FREADME.md。\n\n更有意思的是研究纪律:从 v9 起,每次训练 run 都要先写明预测和失败条件,跑完把结果 append 在后面,不许回头改。当前参考模型是 v19;最新的 v20 实验 169\u002F231,与 v19 的 168 在重训噪声之内,没有取代它。仓库自己统计,v17 recipe 六次重训的标准差是 3.2 个任务——两次单跑差 10 个任务以内都算分不出胜负。这种「没成功的实验也如实留档」的做法,在模型发布里相当少见。\n\n## 所以呢\n\n决策模型卡的位置很具体:agent 工作流里那些「不值得叫一次大模型」的判断——工具该不该现在调、参数是不是用户真说过的、请求该路由给谁、输出过没过护栏。官方给的用法清单就是这些:模型路由、工具选择、参数检查、分流、guardrails、低成本评估,以及混合 agent——难题留给 LLM,机械决策交给 decider,同时降成本降延迟。\n\n赛道也在挤进更多玩家:就在 Strands Decider 开源的同一天,Cloudflare 也发布了开源决策模型 Clef;第三方发布追踪页把 10 月开月概括为「没有前沿大模型,只有一簇小型专用模型」。决策、语音、转录这些「小而专」的品类,正在成为 agent 基建的新一层。\n\nStrands Decider 2B 的差异化不在跑分,而在整条生产线全开放——你可以在自己的 3090 上花 11 小时复现一遍,再改出自己的版本。做 agent 的人值得花一个下午:`pip install strands-decider`,本地起服务,把最贵的判断留给大模型,把其余的交出去。\n\n参考:[Strands Agents 官方博客](https:\u002F\u002Fstrandsagents.com\u002Fblog\u002Fintroducing-strands-decider\u002F)、[GitHub 仓库](https:\u002F\u002Fgithub.com\u002Fstrands-labs\u002Fstrands-decider)。","https:\u002F\u002Fstrandsagents.com\u002Fblog\u002Fintroducing-strands-decider\u002F","033f671e-43d1-4807-82a0-faa1f3725dd9",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"2b4191e8-252e-4281-adbe-fbd2f1839eef","en","Amazon open-sources 2B decision model, full recipe included","Amazon's Strands team open-sourced a 2B agent decision model: one pass, no text output, 115 ms median, full weights, data and scripts all on GitHub.","Amazon's strands-labs team open-sourced Strands Decider 2B on October 1. It writes no code, chats, or summarizes — it only picks from options or rates things on a scale. Weights, training data, and scripts are all on Hugging Face and GitHub under Apache 2.0, running on a local CPU or GPU.\n\n## The category first: what is a decision model\n\nThe wave started with TypeSafe AI's Jev earlier this month. Decision models (also called \"system one\" models) split the labor with LLMs: LLMs generate arbitrary text, while decision models handle three question types — noul (yes\u002Fno), choice (pick one of N), and score (rate on a rubric). In exchange, they are faster at a given size, always answer from the given options, and attach a calibrated confidence to every decision — something the official post stresses that frontier LLM inference APIs do not offer.\n\nThe blog gives a concrete example: given \"Help! My payouts have been failing for 3 days!\" and asked \"Which team should handle this? billing\u002Fsales\u002Fretail\", the model answers billing at confidence 0.768, listing probabilities for all three options.\n\n## Architecture: cut the LM head, add a pointer head\n\nThe approach is surgical: take the Qwen3.5-2B pretrained torso, discard the language-modelling head — the model loses its ability to generate text — and replace it with a pointer head of about a million parameters. The head scores each option by comparing the hidden state at the `\u003Canswer>` position against each option's hidden state. One forward pass, no decoding loop. The torso is adapted with a rank-16 LoRA.\n\nBecause the head holds no per-option parameters, nothing can learn \"the first option is usually right\", and nothing caps how many options a question may carry — label sets are defined by the request, not baked into the weights. All three question types come out of the same masked softmax.\n\nNumbers: on the JevBench public set of 231 tasks, accuracy is 0.723 (167\u002F231); by tier, easy 1.000, standard 0.875, hard 0.505; answers at confidence 0.9 or higher are right about 95% of the time. On latency, an RTX 3090 delivers a 115 ms median and 299 ms 95th percentile; an M3 MacBook handles small tasks in a 153 ms median. Per the third-party board the post cites, it ranks 3rd of 33 in the 2B class, and 1st of 30 when just-over-2B models are excluded.\n\nOne bonus: Qwen3.5-2B-Base is natively multimodal. With `--vision` the server keeps the vision tower, and the same v19 checkpoint answers over images with no retraining — on NaturalBench it is better calibrated (ECE 0.014) than an image-trained 2B decider (0.080).\n\n## The full recipe is open — failures included\n\nThe repository's most unusual part is not the model but the process. Training cost is pressed to a reproducible level: about 11 hours on one RTX 3090 retrains the whole recipe, or 1 hour 10 minutes on eight H100s with FAST mode. Data source licences are listed item by item, and a reproduction contract is stated in the repo.\n\nThe research discipline is more interesting: since v9, every training run states its predictions and failure conditions before training, and the outcome is appended after the run without editing. The reference model today is v19; the latest experiment v20 scored 169\u002F231, within retrain noise of v19's 168, and did not displace it. The repo notes that six retrains of the v17 recipe had a standard deviation of 3.2 tasks — a gap under 10 tasks between single runs is unresolved. Keeping failed experiments honestly on record is rare among model releases.\n\n## So what\n\nDecision models sit at a specific spot: the judgments inside agent workflows that \"don't deserve a frontier LLM call\" — should this tool run now, are these arguments grounded in what the user said, where should this request route, did the output pass the guardrail. The official use-case list is exactly that: model routing, tool selection, argument checking, triage, guardrails, low-cost evaluations, and hybrid agents — LLMs take the hard decisions, deciders take the rote ones, cutting cost and latency together.\n\nThe lane is getting crowded: on the same day Strands Decider shipped, Cloudflare released its own open decision model Clef; a third-party release tracker summed up October's opening as \"no frontier launch, a cluster of small specialist models instead\". Decision, speech, transcription — these \"small and specialized\" categories are becoming a new layer of agent infrastructure.\n\nStrands Decider 2B's differentiation is not the score but the fully open production line — you can retrain it on your own 3090 in 11 hours and fork your own version. If you build agents, spend an afternoon: `pip install strands-decider`, serve it locally, keep the expensive judgments for the big model, and hand the rest over.\n\nReferences: [Strands Agents blog](https:\u002F\u002Fstrandsagents.com\u002Fblog\u002Fintroducing-strands-decider\u002F), [GitHub repository](https:\u002F\u002Fgithub.com\u002Fstrands-labs\u002Fstrands-decider).","amazon-strands-decider-2b","2026-10-04T21:15:00Z","2026-10-04T21:13:03.332066Z","2026-10-04T21:13:03.332074Z",true,"agent",625,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"983fb4d1-6c63-4828-adbf-a59c57e02b64","Perplexity开源决策模型:总分微胜Jev","perplexity-decider-v1-27b-open-source","2026-10-02T21:20:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"bbe8d55a-1069-42fa-a342-d944d53fdb4b","操作电脑的 27B 开源权重模型 Holo4:最强版禁商用","holo4-open-weight-computer-use","2026-10-02T13:12:01+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"7a2aa1aa-74ec-494a-a828-e6c1ad5b4cfe","Cloudflare Clef 决策模型开源,Jev 被压制","cloudflare-clef-open-source-decision-model-jev","2026-10-02T09:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"dc61debd-1d77-4d5a-9d59-5b23c3da07de","蚂蚁开源Realtime-Venus：9B全双工模型边说边干活，三项续聊指标超GPT-4o","ant-realtime-venus-full-duplex-delegation","2026-09-30T23:10:53+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"9103452d-d036-4b0f-bee6-5eda4b7fbf66","Fastino 340M 分类模型:官方基准胜 4B 生成式","fastino-gliner2-5-decide-classifier","2026-09-26T13:09:43+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"0565190a-0bcd-492f-934f-0ad2ab32f485","70万参数2.8MB填一张表:Cua开源CUA-S1,单次前向替代23轮LLM","cua-s1-forms-system-one-model","2026-09-20T13:11:48+00:00"]