[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-microsoft-decision-1-qwen-3-5-9b":3,"topics-all":38,"news-related-d6944f35-e3a6-4d5f-a4e1-a546836f7d75":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},"d6944f35-e3a6-4d5f-a4e1-a546836f7d75","微软基于 Qwen3.5-9B 推出决策模型 Microsoft-Decision-1","微软近日发布面向结构化决策任务的 Microsoft-Decision-1,基座为阿里开源的 Qwen3.5-9B,采用单次推理决策评分机制主打低延迟。微软同时表示后续会把这套决策能力迁移到自研模型与 OpenAI 模型上,逐步替换 Qwen 基座。","## 从聊天助手到决策评分,微软把阿里 Qwen 拉进了结构化决策场景\n\n2026 年 10 月 12 日,微软正式推出 Microsoft-Decision-1,定位是面向结构化决策任务的快速评分模型,基座选用了阿里巴巴开源的 Qwen3.5-9B,通过对基座模型做后训练得到。根据微软官方介绍,这是一套单次推理即可给出决策评分的机制,主打低延迟。\n\n这一动作的关键词不在「又一个决策模型」,而在于「Qwen 成了欧美大厂的现成工作母机」。过去一年,业界已经习惯「中国开源基座 + 海外厂商做领域微调」这种组合,Microsoft-Decision-1 是这条路径上体量最大的一次:微软官方明确表示,未来会用自研模型以及 OpenAI 的模型来逐步替换 Qwen 基座,但当前阶段 Qwen 仍是性能\u002F成本权衡下的最优底座。\n\n## 技术点:Qwen3.5-9B 后训练,核心是「单次推理决策评分」\n\n根据新浪科技、ZAKER 等多家报道,Microsoft-Decision-1 的实现并不是从头训练一个决策模型,而是对 Qwen3.5-9B 做后训练(post-training),重点改造的是推理路径:目标是让模型在一次前向推理中直接输出决策评分,而不是传统的多轮思考或多步工具调用。\n\n这个设计在工程上对应了三个明确收益:\n\n- **延迟**:单次推理的 P50 延迟被压到毫秒级,适合在线决策、实时风控等对响应时间敏感的场景。\n- **成本**:决策评分本质是把复杂问题拆成「打分」而非「生成长文本」,token 消耗远低于通用 LLM 的多步推理。\n- **可控性**:评分输出是结构化数值,可以与传统机器学习流水线(规则、A\u002FB 测试框架)直接对接,而无需再做一层文本解析。\n\n微软在发布材料中给出的一组参考数据是:在延迟维度上,Microsoft-Decision-1 相较 H2O-Lightning-4B 快约 2.5 倍,相较 Jev 1.13.0 快约 2.8 倍;在覆盖文本分类等任务的 36 项基准测试中,准确率达到 83.5%,置信度评分 92.2%。需要注意的是,这些数字来自微软自述,目前尚未见到独立第三方的同口径复现。\n\n## 行业影响:Qwen 不再只是「中文开源选项」\n\nQwen 系列过去两年的故事线,大致可以分成三段:\n\n1. **追平闭源**:Qwen2\u002FQwen3 在多个英文 benchmark 上追平甚至超过同尺寸的 Llama、Mistral,确立「中文系开源也能打」。\n2. **被集成到海外大厂的产品线**:Anthropic、AWS Bedrock 等陆续把 Qwen 作为可选项之一,这是「被列入菜单」阶段。\n3. **被作为底座微调出独立产品**:Microsoft-Decision-1 是这条线上的典型样本,基座 Qwen、上层是微软品牌,产品由微软直接面向客户交付。\n\n第三阶段对阿里和 Qwen 团队的意义在于,基座模型的影响半径不再局限于「下载量」「GitHub Star」,而是开始出现在海外厂商的对外产品页和 SLA 承诺里。\n\n反过来,对微软来说,这条路径也有清晰的商业逻辑:在 Agent、Copilot 等场景下,「决策评分」是高频但低价值密度的子任务,没必要每一次都让 GPT 系列大模型去跑,用一个 9B 量级、几十毫秒内能返回的小模型接住,既能压住成本,也能腾出算力给真正需要长上下文\u002F工具调用的复杂任务。微软在发布时也明确,后续会把这套决策范式迁移到自研模型和 OpenAI 模型上,逐步替换 Qwen 基座——这意味着 Qwen 阶段只是过渡,微软并不会把核心决策能力长期押在外部基座上。\n\n## 所以呢\n\n对关注开源 LLM 的人来说,Microsoft-Decision-1 给出的信号比模型本身更重要:9B 级别的开源基座已经可以承载面向企业级 SLA 的决策类工作负载,而且海外头部厂商愿意为此站台、愿意把它写进产品页。这对 Qwen3.5-9B 而言是一次高质量的「工业级背书」,对整个开源生态则意味着「小模型做专、精模型做决策」的产品形态正在被验证。\n\n下一步值得盯的,是第三方对 Microsoft-Decision-1 在文本分类之外的金融风控、推荐、运维告警等真实业务场景上的复现报告——如果延迟和准确率都能在生产环境里站住,Qwen 系在企业市场的位置还会再上一个台阶。\n\n参考资料:微软官方 Microsoft-Decision-1 发布说明,以及新浪科技、ZAKER 等多家媒体 10 月 12 日的同步报道。","https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85584","d59894d3-308e-4fd8-8865-86dc1eeac4a2",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"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},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"a2e70ce0-a2de-4706-a4ee-6952b45d9fb0","en","Microsoft launches Microsoft-Decision-1 on Qwen3.5-9B","Microsoft has released Microsoft-Decision-1, a fast decision-scoring model aimed at structured decision tasks. The base model is Alibaba's open-source Qwen3.5-9B, and Microsoft applies post-training to deliver single-pass inference with low latency. Microsoft also notes that Qwen will eventually be replaced by its own models and OpenAI models in this role.","## From chat assistant to decision scorer: Microsoft pulls Alibaba's Qwen into structured decisioning\n\nOn October 12, 2026, Microsoft officially released Microsoft-Decision-1, a fast scoring model for structured decision tasks. The base is Alibaba's open-source Qwen3.5-9B, and the model is produced via post-training on that base. According to Microsoft, the design is a single-pass inference decision-scoring mechanism, optimized for low latency.\n\nThe interesting part of this move is not \"yet another decision model\" but the fact that Qwen is now a working foundation for a US big-tech product. Over the past year, the industry has grown used to the \"Chinese open base + overseas fine-tuner\" combo. Microsoft-Decision-1 is the most prominent case of that pattern: Microsoft states that, going forward, it will replace the Qwen base with its own models and OpenAI's models, but at this stage Qwen remains the best cost\u002Fperformance option.\n\n## Technical angle: post-training on Qwen3.5-9B, core idea is single-pass scoring\n\nPer multiple media reports, Microsoft-Decision-1 is not a from-scratch training effort. It is post-training applied to Qwen3.5-9B, with the inference path as the main thing being reworked. The goal is to have the model output a decision score in a single forward pass, rather than the multi-round reasoning or multi-step tool use typical of general LLMs.\n\nThis design maps onto three concrete engineering wins:\n\n- **Latency**: single-pass P50 latency is pushed into the millisecond range, suitable for online decisioning and real-time risk control.\n- **Cost**: decision scoring essentially turns a complex problem into a \"score\" rather than a long generation, so token spend is far below multi-step LLM reasoning.\n- **Controllability**: score outputs are structured numbers that can plug directly into existing ML pipelines (rules, A\u002FB frameworks) without a separate text-parsing layer.\n\nMicrosoft's release materials cite a reference set of numbers: on latency, Microsoft-Decision-1 is roughly 2.5x faster than H2O-Lightning-4B and roughly 2.8x faster than Jev 1.13.0; on 36 benchmark tasks covering text classification and similar problems, it reports 83.5% accuracy and 92.2% confidence. It is worth noting that these figures come from Microsoft itself, and no independent third-party replication on the same setup has been published at the time of writing.\n\n## Industry impact: Qwen is no longer just \"the Chinese open-source option\"\n\nThe Qwen series' story over the past two years roughly breaks into three phases:\n\n1. **Catching up to closed-source**: Qwen2\u002FQwen3 matched or beat same-size Llama and Mistral models on multiple English benchmarks, establishing that \"Chinese-family open source can fight.\"\n2. **Being integrated into overseas big-tech product lines**: Anthropic, AWS Bedrock and others added Qwen as an option — this is the \"on the menu\" phase.\n3. **Being used as a base for independently branded products**: Microsoft-Decision-1 is a textbook example: Qwen as the base, Microsoft as the brand, the product delivered by Microsoft directly to enterprise customers.\n\nFor Alibaba and the Qwen team, phase three matters because the base model's reach no longer stops at \"downloads\" or \"GitHub stars\" — it now appears in overseas vendors' product pages and SLAs.\n\nFor Microsoft, the path also has clear commercial logic. In Agent and Copilot scenarios, \"decision scoring\" is high-frequency but low-value-density — there is no need to send every such call to a GPT-class model. A ~9B model that can return in tens of milliseconds is enough; it keeps cost down and frees up compute for the long-context, tool-using complex tasks. Microsoft also clearly stated that the decision paradigm will be migrated onto its own models and OpenAI's models, gradually replacing the Qwen base — meaning the Qwen phase is a transition, and Microsoft is not betting core decisioning capability on an external base long-term.\n\n## So what\n\nFor people following open-source LLMs, the signal from Microsoft-Decision-1 is more important than the model itself: a 9B-class open base is now able to carry enterprise-SLA decisioning workloads, and an overseas top-tier vendor is willing to put its name on that and write it into a product page. For Qwen3.5-9B this is a high-quality piece of industrial-grade validation; for the broader open-source ecosystem, it confirms that the \"small model for narrow tasks, large model for general reasoning\" product shape is being validated in production.\n\nWhat is worth watching next is third-party reproduction of Microsoft-Decision-1 on real business scenarios beyond text classification — financial risk control, recommendation, ops alerting. If latency and accuracy both hold up in production, Qwen-family models will move up another tier in the enterprise market.\n\nReferences: Microsoft's official Microsoft-Decision-1 release notes, plus sync coverage by Sina Tech, ZAKER and other outlets on October 12.","microsoft-decision-1-qwen-3-5-9b","2026-10-13T03:00:00Z","2026-10-13T03:03:30.174561Z","2026-10-13T03:03:30.174572Z",true,"agent",190,[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},"713ebd28-7abc-4664-acd5-8ba7c6b50d97","扩散 LLM 能否 scale?Mercury 3 用 40 倍规模和 1M 上下文作答","mercury-3-diffusion-scaling","2026-10-19T23:09:33+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"8c6d408b-2913-48b2-bc1f-4e958ad854c5","Space Bunny 谜底揭晓:MiniMax M3.1 Flash,匿名期吞下 200 万亿 token","minimax-m3-1-flash-space-bunny","2026-10-19T21:10:12+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"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":75,"title":76,"news_slug":77,"published_at":78},"bf9bbc74-5720-4829-842e-c3a583a87a8b","Qwen3.8 全模态降价九成逼 Gemini 跟牌","qwen3-8-omni-flash-price-cut","2026-10-04T02:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"18fa7eb9-44e7-4722-ad95-52fc7d795434","Qwen4 开训:阿里把大模型推到 5-10 万亿参数","qwen4-training-5-10-trillion-params","2026-10-03T05:00:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"178858d2-cb75-40e4-b3d1-6ac90add7d96","扩散模型上岗电话客服:推理开口只要320毫秒","mercury-voice-diffusion-llm","2026-09-29T23:07:20+00:00"]