[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-microsoft-mistral-vera-rubin-sovereign":3,"news-related-dfc3dec4-2211-4c7e-b6ff-9e0d9a479ec4":39},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":26,"news_slug":32,"published_at":33,"created_at":34,"modified_at":35,"is_published":36,"publish_type":37,"image_url":13,"view_count":38},"dfc3dec4-2211-4c7e-b6ff-9e0d9a479ec4","微软与 Mistral 签下数十亿美元协议:Vera Rubin GPU 上的「欧洲主权云」开始落地","\n**微软与 Mistral 7 月 21 日联合宣布的扩展合作,把欧洲 AI 基础设施从概念推进到工程现实。** 这次合作的核心不是单一模型,而是三个变量同时落地:数千颗 NVIDIA Vera Rubin GPU 进入 Mistral 欧洲数据中心;Mistral Medium 3.5 与 OCR 4 同步接入 Microsoft Foundry 与 Copilot Studio;Azure 与 Azure Local 打通\"云端 + 云连接 + 完全离线\"三种部署形态。\n\n**Mistral Medium 3.5 是这次唯一开权重的旗舰——128B 密集架构,256k 上下文,SWE-Bench Verified 跑出 77.6%。** 更值得注意的工程信号是:它支持\"少至 4 张 GPU 自托管\",并把视觉编码器从零重训以支持任意长宽比图片,直接对位 Devstral 2 与 Qwen3.5-397B-A17B。OCR 4 是 Document AI 栈的新骨干,主打\"段落级 bounding box + 结构化 block label\",按页计价 $4\u002F1000、标注页 $5\u002F1000,在金融文档、合同评审这类需要坐标级可解释的 RAG 场景里,直接成为 Mistral 进 Foundry 的入口产品。\n\n**算力底座是 Vera Rubin NVL72,不是上一代 Blackwell。** NVIDIA 把这次部署定性为\"Agentic AI 驱动的能源高效基础设施\",背后是 Vera Rubin 在每兆瓦 token 吞吐上的代际优势——把欧洲主权云的总拥有成本和延迟,推到与美国一线云厂商可比的区间。对 Mistral 而言,这是用\"算力主权 + 模型主权\"两条腿站住欧盟市场;对微软而言,Sovereign Cloud 阵营多了一张能正面应对 Google 与 AWS 区域 AI 服务的欧洲牌。\n\n**对开发者的实际意义是 Mistral 中端模型第一次有了 Azure SLA。** 之前要在企业内网用 Mistral,只能私有化部署或直连 La Plateforme;现在 Foundry + Foundry Local + Azure Local 三件套,意味着同一个 Medium 3.5 既能在公有云弹性调用,也能在工厂车间、医院内网、政府隔离网跑同一套 API——这是 Mistral 进入大型政企采购名单的硬性门槛,也是开源权重模型与 OpenAI、Anthropic 在企业市场抢位置的真正切入点。","https:\u002F\u002Fnews.microsoft.com\u002Fsource\u002F2026\u002F07\u002F21\u002Fmicrosoft-and-mistral-expand-strategic-partnership-to-give-enterprises-and-regulated-industries-frontier-ai-they-can-control\u002F","9f1ed564-b7ac-4084-b1a8-4d129772db34",[10,14,17,20,23],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":21,"name":22,"slug":22,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":24,"name":25,"slug":25,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[27],{"id":28,"lang":29,"title":30,"summary":31,"content":31},"01e927f1-7e1d-402b-9ede-99672f037664","en","Microsoft-Mistral deal: a European sovereign cloud on Vera Rubin","The expanded partnership Microsoft and Mistral jointly announced on July 21 pushes European AI infrastructure from concept to engineering reality. The deal's core is not a single model, but three variables landing at once: thousands of NVIDIA Vera Rubin GPUs entering Mistral's European data centers; Mistral Medium 3.5 and OCR 4 being plugged into Microsoft Foundry and Copilot Studio; and Azure + Azure Local providing three deployment shapes — cloud, cloud-connected, and fully offline. Mistral Medium 3.5 is the only flagship with open weights in this round — a 128B dense architecture, 256k context, scoring 77.6% on SWE-Bench Verified. The more interesting engineering signal is that it supports \"self-hosting on as few as 4 GPUs\", and the vision encoder is retrained from scratch to support arbitrary-aspect-ratio images, directly targeting Devstral 2 and Qwen3.5-397B-A17B. OCR 4 is the new backbone of the Document AI stack, with paragraph-level bounding boxes and structured block labels, priced at $4\u002F1000 pages or $5\u002F1000 annotated pages, becoming Mistral's entry product into Foundry for financial documents, contract review and other RAG scenarios requiring coordinate-level interpretability. The compute foundation is Vera Rubin NVL72, not last-gen Blackwell. NVIDIA positions this deployment as \"Agentic-AI-driven energy-efficient infrastructure\", backed by Vera Rubin's generational advantage in tokens-per-megawatt — pushing the European sovereign cloud's total cost of ownership and latency into a range comparable to top-tier U.S. clouds. For Mistral, this is \"compute sovereignty + model sovereignty\" standing on two legs in the EU market; for Microsoft, the Sovereign Cloud camp gets a European card that can stand up to Google and AWS regional AI services head-on. For developers, the practical significance is that Mistral's mid-tier model has Azure SLA support for the first time. Previously, to use Mistral inside an enterprise intranet you could only self-host or connect directly to La Plateforme; now the Foundry + Foundry Local + Azure Local trio means the same Medium 3.5 can be invoked elastically on the public cloud, or run with the same API on a factory floor, in a hospital intranet, or in a government-isolated network — the hard threshold for Mistral to enter large-enterprise and government procurement lists, and the real wedge for open-weight models to compete with OpenAI and Anthropic for enterprise market share.","microsoft-mistral-vera-rubin-sovereign","2026-07-22T02:00:00Z","2026-07-22T06:10:58.277318Z","2026-08-19T02:08:40.142862Z",true,"agent",92,{"items":40},[41,46,51,56,61,66],{"id":42,"title":43,"news_slug":44,"published_at":45},"40095b51-97b0-4fd4-9b1d-f636c970572e","阿里 Qwen 团队发布 Qwen3.8-Max:2.4 万亿参数 MoE 模型首度开放权重","qwen3-8-max-2-4t-moe-open-weights","2026-08-07T02:00:00+00:00",{"id":47,"title":48,"news_slug":49,"published_at":50},"804b44fb-66c6-4355-83a4-b3a03a776d2a","Inkling-Small 开放权重：12B 激活参数换来更高 Agent 效率，也暴露事实性短板","inkling-small-multimodal-moe-efficiency","2026-08-05T16:32:13+00:00",{"id":52,"title":53,"news_slug":54,"published_at":55},"a151db0c-d832-4df2-ac03-2d4e58b26e99","Kimi K3 跑通 MiniTriton:Moonshot 让 LLM 第一次从零编译出自己的 GPU 编译器","kimi-k3-minitriton-gpu-compiler","2026-07-26T14:00:00+00:00",{"id":57,"title":58,"news_slug":59,"published_at":60},"5bfdf32b-44eb-4eb5-a98b-39e921168182","九天内连发五款前沿模型:7 月的大模型军备赛,真正决胜负的不再是 benchmark","july-2026-five-frontier-models","2026-07-23T12:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"cf01282f-8a64-49a8-a608-9b806ccfbea3","Mira Murati 实验室 Inkling 开源：975B MoE 不卷\"最强\"，押注\"可定制\"","thinking-machines-inkling","2026-07-15T22:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"1a27bedc-012d-4d63-85e1-ddc57aabd8bf","ByteDance UniVR 让模型「在视觉空间里思考」：34B 参数逼近 Gemini 3 Pro + Nano Banana 2","bytedance-univr-34b","2026-07-14T12:10:00+00:00"]