[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-microsoft-decision-1-rebase-qwen-to-mai-openai":3,"topics-all":41,"news-related-ff3ccd81-4ff3-468e-966c-21ef57126883":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"ff3ccd81-4ff3-468e-966c-21ef57126883","微软拍板:Decision-1 之后,会把 Qwen 换成 MAI 和 OpenAI","微软在 Microsoft-Decision-1 一文中第一次明示,这个基于 Qwen3.5-9B 后训练出的决策模型,后续会重新基座在自家 MAI 和 OpenAI 模型上。从\"借中国开放权重试水\"过渡到\"自研 + 联盟接力\",决策模型赛道被巨头正式收编。","微软在 10 月 9 日推出 Microsoft-Decision-1 时,把一句关键承诺藏在 How it works 中段:\n\n> To build Microsoft-Decision-1, we post trained Qwen3.5-9B for fast, single-pass decision scoring and will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI.\n\n2.5 倍领先 H2O-Lightning-4B、35 倍领先 GPT-6 Sol、36 项基准 83.5% 准确率这些数字,是背景;更值得看的是这句 rebasing 承诺。来源:https:\u002F\u002Fcommandline.microsoft.com\u002Fmicrosoft-decision-1-model-foundry\u002F\n\n## 决策模型不是 LLM\n\n决策模型接收固定选项、返回校准概率、不生成自由文本——天然适合路由、护栏、AI 评判这三种 agent 工作流里最费钱的环节。Decision-1 技术选型是后训练基座 Qwen3.5-9B(9B 推理质量够、单卡延迟低),任务支持 yes\u002Fno、多选、量表打分,概率就是 API 输出,不做置信度\"再来一次\"。同请求做 8 种扰动,平均决策翻转仅 1.3%,选项顺序反转、改写、重排时直接 0 翻转。整体结构不在 LLM 能力范围内,只是借 Qwen3.5-9B 的语言理解底子,把它压成\"只挑选项\"的窄任务头。\n\n## 微软自家化:MAI 和 OpenAI 接棒\n\nJev 路线是\"后训练一个开源底座 + 收窄任务头\"。微软先用 Qwen3.5-9B 跑通流水线(输出免费、输入每百万 token 0.042 美元、文本分类成本不到 GPT-6 Sol 的 1\u002F20),再把基座切到自研——MAI 是自研系列代号,Decision-1 是它第一次以\"接替 Qwen\"姿态露脸,OpenAI 这里是过渡缓冲而非\"把 OpenAI 模型拿来跑决策\",首发保留 Qwen3.5-9B,等 MAI 到位再切换。基座替换本身不稀奇,但 Decision-1 的特殊在于:决策模型这个原本\"快速开源 + 社区拼跑分\"的赛道,被一家没出过决策模型的厂用\"先借中国开放权重跑通、再换自家模型替换\"的策略正式收编。\n\n## 对中国开放权重的样板\n\n微软用的是 post trained,不是 fine-tuned。Qwen3.5-9B 在这条线上的角色是\"加速器\"不是\"基础设施\"——Microsoft AI 的 MAI 系列此前多次从 Llama \u002F Mistral 转自研,Decision-1 是同一剧本搬到决策模型赛道。**被采纳并不等于被依赖;首发版本最快被换掉,反而比\"长期卡在首发版本\"更说明问题——首发版只能\"跑通管道\",真正决定长期方向的是厂商自己的底座。**\n\n10 月 9 日这条公告里最值得读的,不是 35 倍速也不是 1.3% 翻转率,而是那句 will soon rebase it on other models——这是大厂把决策模型从 open-source 路线收编到 first-party 路线的明确信号。","https:\u002F\u002Fcommandline.microsoft.com\u002Fmicrosoft-decision-1-model-foundry","144f3573-80d9-4d33-b645-ea97eabc398b",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"c33b1bbc-d6ce-4f61-9d5d-1a0704a6a09b","ai-policy",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"42e59a88-7795-47dc-a334-ef1e72c24347","openai",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",{"id":25,"name":26,"slug":26,"description":14,"color":14},"95d7995a-fddb-47ba-b8e6-e976ac65414b","strategy",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"1492fd58-00e4-4c67-9e75-0c9470e6a75a","en","Microsoft Decision-1 will rebase off Qwen onto MAI and OpenAI","Microsoft-Decision-1 launches on Qwen3.5-9B and then rebases on MAI and OpenAI. Decision-model lane consolidates from open-source to first-party.","When Microsoft introduced Microsoft-Decision-1 on October 9, the most consequential sentence sat in the middle of the How it works section, not the headline benchmark table:\n\n> To build Microsoft-Decision-1, we post trained Qwen3.5-9B for fast, single-pass decision scoring and will soon rebase it on other models, including Microsoft AI (MAI) and OpenAI.\n\nThe lineup behind it (Cloudflare Clef, Perplexity pplx-decider, GLM-5.3-Flash System One, firelex Jeff) confirms that the decision-model category is moving from open-source baseline to a tier-one vendor consolidation. Source: https:\u002F\u002Fcommandline.microsoft.com\u002Fmicrosoft-decision-1-model-foundry\u002F\n\n## What Decision-1 actually is\n\nDecision models take a fixed option set, return a calibrated probability per option, and never generate free text. Routing, guardrails, and AI-as-judge are the three workloads where they save the most latency and cost. Decision-1 ships in Microsoft Foundry and OpenRouter:\n\n- Post-training base: Qwen3.5-9B (Alibaba open-weights 9B)\n- Tasks supported: yes\u002Fno, multi-choice, rating, rubric-based grading\n- Calibration: the probability is the API output, not a ranking\n- Robustness check: 8 perturbation variants of the same prompt, average flip rate 1.3%, 0% flips when option order is shuffled, reversed, or paraphrased\n- Safety: 5,250 requests across 11 safety benchmarks, refused harmful content while keeping utility high\n\nMicrosoft headline is the latency gap: 2.5x faster than H2O-Lightning-4B (runner-up) and 35x faster than GPT-6 Sol. 36-benchmark accuracy tops the table at 83.5% across roughly 150,000 questions blinded from training; calibration score 92.2%, second only to Quyet-1.0-Large. Pricing: input 0.042 USD per million tokens, output free, text-classification cost under one-twentieth of GPT-6 Sol.\n\n## Why the rebasing plan matters\n\nPost-training on an open-weight base is the standard playbook for any team entering a narrow task lane without committing to a full base-model training run. What is unusual is that Microsoft chose to publish the rebasing roadmap up front. Three signals to read from that:\n\n1. MAI is the long-term base. Microsoft AI s MAI line has been referenced repeatedly over the past year but has not shipped as a flagship. Decision-1 is the first place MAI appears as a successor to an existing public model. The market now has a clearer read on where MAI general-purpose capability is being readied: not in chat, but in narrow inference workloads where calibration matters more than conversational fluency.\n2. OpenAI is a transitional safety net. The phrase including OpenAI next to MAI is not we will move to OpenAI models for decisioning. GPT-6 Sol has just been outpaced by Decision-1 on the same benchmarks; OpenAI here is the fallback for the period between Qwen3.5-9B and MAI readiness, not the destination.\n3. Qwen3.5-9B is intentionally the launch base. Microsoft could have waited; it chose not to. Launching with Qwen3.5-9B lets Decision-1 compete on benchmark tables today, where every week of delay is a week Cloudflare, Perplexity, and zhipu (GLM-5.3-Flash System One) extend their lead. The roadmap to rebasing is essentially: we win this week, we rebase next quarter.\n\n## The decision-model lane is now crowded\n\nBetween late September and October 9, five first-party decision-model releases have shipped:\n\n| Model | Date | Base |\n|---|---|---|\n| System One (GLM-5.3-Flash) | 2026-09-28 | Zhipu GLM-5.3-Flash |\n| Jeff (firelex) | 2026-09-29 | Self-built 2B |\n| Perplexity pplx-decider-v1-27b | 2026-10-02 | Self-built 27B |\n| Cloudflare Clef | 2026-10-02 | Self-built |\n| Microsoft-Decision-1 | 2026-10-09 | Qwen3.5-9B (rebase to MAI \u002F OpenAI) |\n\nJev, the model that defined this category in H1 2026, is now flanked from every side: big-tech entrants (Microsoft, Cloudflare), strong-startup entrants (Perplexity), open-weights repos (Jeff), and incumbent-derivative entries (GLM-5.3-Flash System One). The crowded lane is the precondition that makes the first-party consolidation play rational.\n\n## What this signals for Chinese open weights\n\nThe standard read on Microsoft picks Qwen3.5-9B is an open-weight Chinese model gets adopted by a US hyperscaler. The more precise read is the rebasing sentence: Qwen3.5-9B is the launch vehicle, not the long-term base. Three downstream effects:\n\n- Adoption is real, but bounded. Microsoft text-classification cost-claim (one-twentieth of GPT-6 Sol) and API pricing (output free) only make economic sense if the base model is small enough to serve on commodity hardware. Qwen3.5-9B delivers exactly that. But the same paragraph that publishes the cost numbers also commits to a rebasing. So Qwen3.5-9B absorbs the launch workload and gets replaced when MAI is ready.\n- The pattern is the playbook, not the partnership. Microsoft AI MAI line has previously transition-tested on Llama and Mistral bases; Decision-1 is the same script in a new lane. The decision-model race is therefore not which open base will Microsoft standardize on, it is how fast can Microsoft ship a workable first-party base.\n- The signal for other tier-one decision-model entrants. Cloudflare, Perplexity, and Zhipu are all running their own self-built bases; only Microsoft is openly running on a third-party Chinese open-weight base for launch. The implicit bet is that the Qwen3.5-9B-based Decision-1 is good enough for the launch window but not good enough for the MAI-window; whether this bet pays off depends on MAI decision-scoring capability landing in Q1 2027.\n\nThe cleanest way to read the announcement is to treat the rebasing sentence as the real headline. Microsoft shipped a Qwen-based decision model on October 9, but the more important sentence is the one saying they will not stay on Qwen3.5-9B for long.","microsoft-decision-1-rebase-qwen-to-mai-openai","2026-10-12T09:00:00Z","2026-10-12T09:15:12.321555Z","2026-10-12T09:15:12.321572Z",true,"agent",152,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"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":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"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":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"d83c4910-6dce-4171-8a44-9f68c9f841af","Qwen-Audio-3.1-ASR开源:转写自带说话人","qwen-audio-3-1-asr-speaker-diarization","2026-10-22T17:13:29+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"405b8f47-8862-4684-ab0e-d552ce093072","MJev-Spatial:空间理解不生成文字,单次前向快5.6倍","mjev-spatial-classification-spatial-intelligence","2026-10-18T17:14:25+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"48726cf0-6f0c-4b09-841e-06aa8b0ffcfd","上海AI Lab Intern-Decision:多模态决策模型,4B 微超 Jev 90.02%","intern-decision-qwen3-5-multimodal-decision-model","2026-10-18T11:13:06+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"d6944f35-e3a6-4d5f-a4e1-a546836f7d75","微软基于 Qwen3.5-9B 推出决策模型 Microsoft-Decision-1","microsoft-decision-1-qwen-3-5-9b","2026-10-13T03:00:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"59d14de3-0e17-4202-8e7d-ad0bc51e3471","Qwen-Image-2.1-Turbo开源:8步去噪出图","qwen-image-2-1-turbo-8-step","2026-10-11T15:05:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"40fc3d8d-da30-43cb-a945-79b9c34fbb09","微软决策模型Decision-1:比Sol快35倍,输出免费","microsoft-decision-1-fast-decision-model","2026-10-10T15:10:00+00:00"]