跳到主要內容

MAI-Base-2評測:微軟2.8兆參數模型,跑分未公開 | MAI-Base-2 Review: 2.8T Params, No Benchmarks Yet

By Kit 小克 | AI Tool Observer | 2026-10-16

🇹🇼 MAI-Base-2評測:微軟2.8兆參數模型,跑分未公開

MAI-Base-2是什麼:微軟2.8兆參數基礎模型重磅登場

微軟在2026年10月16日宣布完成新一代基礎模型MAI-Base-2的預訓練,總參數高達2.8兆,採用混合專家(MoE)架構,實際運算時只啟動800億參數。微軟宣稱這顆模型的預訓練運算效率比DeepSeek-V4-Pro-Base高出4到12倍,是這幾天AI圈討論度最高的話題之一。

規格看起來真的猛

  • 總參數2.8兆,啟動參數800億,比上一代MAI-Base-1大上不少
  • 訓練資料量達50兆token,其中18%是圖文混合資料,代表這顆模型從預訓練階段就支援多模態
  • 用16,000張GB200 GPU跑了38天訓練完成
  • 知識截止日為2026年3月
  • 微軟強調訓練資料是「有機」取得、沒有用其他模型蒸餾——這點其實也跟它未來打算透過Foundry授權權重給夥伴有關,商業考量跟技術說法摻在一起

跑分呢?其實還沒有

這才是MAI-Base-2目前最該被劃重點的地方:微軟只公開了預訓練階段的內部評測結果,開發者真正在意的程式碼能力、Agent任務、數學推理、電腦操作等後訓練(post-training)跑分完全沒有公布。而且拿來證明效率提升的指標是BPB(bits-per-byte),這只能說明模型「預測文字的準確度」,不代表它能不能真的把問題解出來。更麻煩的是,這些評測集都是微軟內部的,外界沒辦法重現或驗證這些數字。

目前也沒有公開發布日期、沒有定價,微軟只說會透過Foundry把模型權重授權給「受信任的夥伴」,一般開發者短期內摸不到這顆模型。

現在該怎麼看這則新聞

如果你是開發者或企業IT主管,MAI-Base-2目前仍停留在「預訓練完成、數字漂亮」階段,跟一顆真正能拿來跑生產環境的模型中間還有後訓練、安全對齊、實際上線測試等好幾個關卡。微軟過去也不是沒發生過跑分亮眼但實際體感普通的案例,在看到真正的coding/agent benchmark跟公開API之前,這些效率數字先保留一點懷疑比較健康。

好不好用,試了才知道。


🇺🇸 MAI-Base-2 Review: 2.8T Params, No Benchmarks Yet

What Is MAI-Base-2: Microsoft's 2.8-Trillion-Parameter Foundation Model

On October 16, 2026, Microsoft announced it has finished pre-training MAI-Base-2, its next flagship foundation model. The model has 2.8 trillion total parameters in a mixture-of-experts (MoE) architecture, with only 80 billion active at inference time. Microsoft claims MAI-Base-2 delivers 4-12x better pre-training FLOPs-efficiency than DeepSeek-V4-Pro-Base, which is why it's dominating AI news feeds this week.

The Spec Sheet Looks Impressive

  • 2.8 trillion total parameters, 80B active, a significant jump from MAI-Base-1
  • Trained on 50 trillion tokens, with 18% image-text data, making it multimodal from the pre-training stage onward
  • Trained over 38 days on 16,000 GB200 GPUs
  • Knowledge cutoff: March 2026
  • Microsoft says the training data is "organic," not distilled from other models, a claim that doubles as a commercial pitch since it underpins Microsoft's plan to license the weights to partners through Foundry

Benchmarks? Not Yet.

This is the part that matters most about MAI-Base-2 right now: Microsoft has only published internal pre-training evaluation numbers. The post-training benchmarks developers actually care about, coding, agentic tasks, math reasoning, computer use, simply aren't out yet. The headline efficiency claim is based on BPB (bits-per-byte), a metric that measures how well a model predicts text, not whether it can actually solve a problem. And since the eval sets are internal, there's no way for outside researchers to reproduce the numbers.

There's also no release date and no pricing. Microsoft says it will license model weights to "trusted partners" through Foundry, meaning most developers won't get hands-on access anytime soon.

How to Read This News Right Now

For developers and engineering leads, MAI-Base-2 is currently at the "pre-training done, numbers look great" stage, several steps away from a model you can actually run in production, including post-training, safety alignment, and real-world testing. Microsoft has shipped impressive-looking benchmarks before that didn't fully translate to real-world feel. Until independent coding and agent benchmarks, plus a public API, show up, it's worth holding the efficiency claims at arm's length.

好不好用,試了才知道 — you won't know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles


AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

留言

這個網誌中的熱門文章

Google Ironwood TPU v7 推理專用晶片解析:效能追平 NVIDIA、成本低 44%,AI 晶片戰爭正式開打 | Google Ironwood TPU v7 Explained: Matching NVIDIA Performance at 44% Lower Cost — The AI Chip War Heats Up

Claude Code 實測:AI 幫你寫程式到底行不行? | Claude Code Review: Can AI Really Code for You?

Cursor vs GitHub Copilot vs Claude Code:AI 程式助手大比拼 | AI Coding Assistants Compared: Cursor vs GitHub Copilot vs Claude Code