[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-microsoft-bets-on-controllable-domain-tuned-models-zh":3,"article-related-microsoft-bets-on-controllable-domain-tuned-models-zh":31,"series-model-release-f26334ab-dd8b-49c2-a49e-7fc376200f2b":83},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":23,"views":27,"created_at":28,"published_at":29,"topic_cluster_id":30},"f26334ab-dd8b-49c2-a49e-7fc376200f2b","microsoft-bets-on-controllable-domain-tuned-models-zh","微軟押注可控、領域調校模型，而不是更大的通用模型","\u003Cp data-speakable=\"summary\">微軟正在把 AI 堆疊的重心，從通用大模型轉向可控、領域調校的小型模型。\u003C\u002Fp>\u003Cp>我認為微軟這次押對了方向：企業 AI 的勝負，不在誰能做出最會\u003Ca href=\"\u002Fnews\u002Fchatgpt-grew-from-chatbot-to-platform-zh\">聊天\u003C\u002Fa>的模型，而在誰能把模型塞進真實工作流、控制成本，並保住資料與治理邊界。Build \u003Ca href=\"\u002Fnews\u002Fmicrosoft-build-2026-agents-into-systems-zh\">2026\u003C\u002Fa> 的 MAI 佈局很直接地證明了這點：MAI-Image-2.5 進入 PowerPoint 與 OneDrive，MAI-Transcribe-1.5 串進 \u003Ca href=\"\u002Ftag\u002Fcopilot\">Copilot\u003C\u002Fa>、Teams、\u003Ca href=\"\u002Ftag\u002Fgithub\">GitHub\u003C\u002Fa> 和 Dynamics 365 Contact Centre，MAI-\u003Ca href=\"\u002Fnews\u002Ffable-5-claude-code-like-coworker-zh\">Code\u003C\u002Fa>-1-Flash 甚至成為 \u003Ca href=\"\u002Ftag\u002Fvs-code\">VS Code\u003C\u002Fa> 的預設選項。這不是研究展示，而是把模型當作基礎設施來鋪。\u003C\u002Fp>\u003Ch2>第一個論點\u003C\u002Fh2>\u003Cp>微軟優先考慮的是部署效率，而不是模型奇觀。MAI-Code-1-Flash 只有 5B 參數，卻在 SWE Bench Pro 拿到 51%；同時，MAI-Thinking-1 以 35B active parameters 的 MoE 架構與 256K context window 進場。這組數字傳達的訊息很清楚：企業不需要最大模型，需要的是足夠好的模型，能在延遲、成本與品質之間取得穩定平衡。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781331468190-ymfp.png\" alt=\"微軟押注可控、領域調校模型，而不是更大的通用模型\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>更重要的是，這些模型不是孤立存在。它們被放進使用者已經在工作的地方，像是 PowerPoint、Teams、GitHub 和 Dynamics 365。這種分發方式比單純推出一個聊天頁面更有價值，因為它直接縮短了從「能用」到「真的被用」的距離。對企業來說，AI 的採納率往往不是被能力限制，而是被整合摩擦拖垮。\u003C\u002Fp>\u003Ch2>第二個論點\u003C\u002Fh2>\u003Cp>微軟真正想建立的，是一個把模型、工具、晶片與通路綁在一起的平台優勢。Build 2026 把 MAI-Thinking-1 與 Maia 200 綁定，並宣稱再帶來 1.4x performance-per-watt 改善，同時又把模型開放到 Foundry、OpenRouter、Fireworks 和 Baseten。這代表它不是只賣模型，而是在賣一個可被多路接入、可被放大部署的 AI 堆疊。\u003C\u002Fp>\u003Cp>這種策略的核心，是把可控性變成護城河。Frontier Tuning 與 \u003Ca href=\"\u002Ftag\u002Freinforcement-learning\">reinforcement learning\u003C\u002Fa> environments 的設計，明確指向企業可以用自己的資料建立任務專用 agent，而且控制權留在客戶邊界內。對多數企業而言，這比一個通用聊天介面更有說服力，因為它把 AI 從「租來的能力」變成「可管理的資產」，也更符合合規與內控需求。\u003C\u002Fp>\u003Ch2>反方可能怎麼說\u003C\u002Fh2>\u003Cp>反對者會說，這仍然是 \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> 戲法，只是包裝得更像企業產品。微軟大量引用 97% 的 AIME 25、53% 的 SWE Bench Pro，以及影像編修與轉錄排行榜，聽起來仍像是在追逐模型分數，而不是解決業務問題。換句話說，它只是把通用模型軍備競賽，改寫成多模型、可路由的版本。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781331468532-3s99.png\" alt=\"微軟押注可控、領域調校模型，而不是更大的通用模型\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>另一個質疑也很合理：模型越多，管理越複雜。影像、語音、轉錄、思考、寫碼各一套，對平台方是產品線，對企業方卻可能是選型、維運與治理負擔。若抽象層做得不好，所謂的可控性就會變成新的整合成本，最後只是把複雜度從模型內部搬到系統外部。\u003C\u002Fp>\u003Cp>但這個批評只對了一半。微軟不是在說 benchmark 本身就是產品，而是在說 benchmark 證明這些領域模型已經足夠好、足夠快、也足夠便宜，可以取代許多預設的通用模型工作負載。至於多模型帶來的複雜度，微軟的答案就是平台層：Foundry、Copilot、Teams、VS Code 與 Azure 共同承擔編排。若這層做得好，複雜性會被隱藏，而不是被放大。\u003C\u002Fp>\u003Ch2>你能做什麼\u003C\u002Fh2>\u003Cp>如果你是工程師，不要再把模型選型當成品牌偏好，而要當成架構決策：先定義任務、資料邊界、延遲目標與成本上限，再決定要用通用模型、領域模型，還是調校後的小模型。如果你是 PM 或創辦人，重點更直接：優先設計可控的工作流，而不是追逐最炫的 demo。真正能落地的 AI，不是回答最漂亮的那個，而是能長期嵌進你的流程、數據與治理制度的那個。\u003C\u002Fp>","微軟正在把 AI 堆疊的重心，從通用大模型轉向可控、領域調校的小型模型，因為真正的企業價值來自部署、成本與治理，而不是 demo 的震撼感。","microsoft.ai","https:\u002F\u002Fmicrosoft.ai\u002Fnews\u002Fmicrosoft-build-2026-mai-keynote-transcript\u002F",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781331468190-ymfp.png","model-release","zh","0a587641-eb12-4267-8c2f-66552da4971f",[17,18,19,20,21,22],"Microsoft","AI stack","domain-tuned models","controllability","enterprise AI","platform strategy",[24,25,26],"微軟押注的是可部署、可控、可治理的領域模型，而不是單一通用大模型。","模型優勢正在被平台、晶片與分發通路放大，形成企業級 AI 堆疊。","對工程師與產品團隊來說，AI 選型應以工作流、成本與控制權為準。",0,"2026-06-13T06:17:20.311904+00:00","2026-06-13T06:17:20.303+00:00","0ccb5d2e-69f1-4354-a3e0-cb370221cd95",{"tags":32,"relatedLang":42,"relatedPosts":46},[33,34,36,38,40],{"name":20,"slug":20},{"name":17,"slug":35},"microsoft",{"name":18,"slug":37},"ai-stack",{"name":21,"slug":39},"enterprise-ai",{"name":19,"slug":41},"domain-tuned-models",{"id":15,"slug":43,"title":44,"language":45},"microsoft-bets-on-controllable-domain-tuned-models-en","Microsoft is betting the AI stack on controllable, domain-tuned models","en",[47,53,59,65,71,77],{"id":48,"slug":49,"title":50,"cover_image":51,"image_url":51,"created_at":52,"category":13},"0bb91791-f4b6-4d51-899c-6eeb239f942a","microsoft-mai-models-build-2026-zh","Microsoft把 Copilot 拉回主場","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781330585064-s9ya.png","2026-06-13T06:02:35.160901+00:00",{"id":54,"slug":55,"title":56,"cover_image":57,"image_url":57,"created_at":58,"category":13},"eaafd0fe-cb56-40cd-80f8-c203e3d72f03","gpt-5-4-thinkng-pro-mini-nano-release-zh","GPT-5.4 率先登場，mini、nano 跟進","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781326973205-xufk.png","2026-06-13T05:02:19.576989+00:00",{"id":60,"slug":61,"title":62,"cover_image":63,"image_url":63,"created_at":64,"category":13},"09fe28b5-aae5-4bac-b3bd-9a261e4c99a1","mimo-v2-flash-openrouter-benchmarks-pricing-zh","MiMo-V2-Flash 直衝開源 SWE-bench","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781321565467-96el.png","2026-06-13T03:32:17.367685+00:00",{"id":66,"slug":67,"title":68,"cover_image":69,"image_url":69,"created_at":70,"category":13},"0a9dbc64-2e51-494d-b6b6-21ecfd8dd1f5","minimax-m3-1m-token-coding-power-zh","MiniMax M3 把 1M Token 送進寫碼場","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781295477857-78hl.png","2026-06-12T20:17:28.037784+00:00",{"id":72,"slug":73,"title":74,"cover_image":75,"image_url":75,"created_at":76,"category":13},"8ca34749-2efa-4b24-b2bd-c0fb66062b49","openai-confidential-ipo-us-stock-market-zh","OpenAI 低調送件 IPO","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781294579049-ltqt.png","2026-06-12T20:02:33.419196+00:00",{"id":78,"slug":79,"title":80,"cover_image":81,"image_url":81,"created_at":82,"category":13},"b15b0887-bd5b-43e6-ac42-23939d0f4e92","google-gemini-35-pro-june-2m-token-launch-zh","Gemini 3.5 Pro 6月登場，2M Token 夠猛","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1781204585839-bdsh.png","2026-06-11T19:02:36.371587+00:00",[84,89,94,99,104,109,114,119,124,129],{"id":85,"slug":86,"title":87,"created_at":88},"58b64033-7eb6-49b9-9aab-01cf8ae1b2f2","nvidia-rubin-six-chips-one-ai-supercomputer-zh","NVIDIA Rubin 把六顆晶片塞進 AI 機櫃","2026-03-26T07:18:45.861277+00:00",{"id":90,"slug":91,"title":92,"created_at":93},"0dcc2c61-c2a6-480d-adb8-dd225fc68914","march-2026-ai-model-news-what-mattered-zh","2026 年 3 月 AI 模型新聞重點","2026-03-26T07:32:08.386348+00:00",{"id":95,"slug":96,"title":97,"created_at":98},"214ab08b-5ce5-4b5c-8b72-47619d8675dd","why-small-models-are-winning-on-device-ai-zh","小模型為何吃下裝置端 AI","2026-03-26T07:36:30.488966+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"785624b2-0355-4b82-adc3-de5e45eecd88","midjourney-v8-faster-images-higher-costs-zh","Midjourney V8 變快了，也變貴了","2026-03-26T07:52:03.562971+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"cda76b92-d209-4134-86c1-a60f5bc7b128","xiaomi-mimo-trio-agents-robots-voice-zh","小米 MiMo 三模型瞄準代理、機器人與語音","2026-03-28T03:05:08.779489+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"9e1044b4-946d-47fe-9e2a-c2ee032e1164","xiaomi-mimo-v2-pro-1t-moe-agents-zh","小米 MiMo-V2-Pro 登場：1T MoE 模型","2026-03-28T03:06:19.002353+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"c4b6186f-bd84-4598-997e-c6e31d543c0d","cursor-composer-2-agentic-coding-model-zh","Cursor Composer 2 走向代理式寫碼","2026-03-28T03:13:06.422716+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"e112e76f-ec3b-408f-810e-e93ae21a888a","apple-siri-gemini-distilled-models-zh","Apple Siri 牽手 Gemini 的真相","2026-03-29T04:52:57.886544+00:00",{"id":125,"slug":126,"title":127,"created_at":128},"c679b51f-194a-463b-87fc-7695256ff752","mimo-v2-pro-vs-omni-vs-flash-2026-zh","MiMo V2 Pro、Omni、Flash 怎麼選","2026-04-02T01:18:43.576128+00:00",{"id":130,"slug":131,"title":132,"created_at":133},"3b988fd7-6749-4f01-ba25-c0ad7486dc31","z-ai-glm-5v-turbo-design2code-claude-zh","GLM-5V-Turbo 在 Design2Code 贏了…","2026-04-02T04:03:36.31741+00:00"]