Microsoft AI finishes pre-training MAI-Base-2: 2.8T parameters, 80B active, trained in-house on 16k GB200s in 38 days

Microsoft AI says it has finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion parameters in total and 80 billion active per token. It is the base model, not a product: Microsoft says it is now being post-trained, and that it will share results "soon". The announcement came October 15 in a technical post on microsoft.ai.

THE MODEL

2.8T total / 80B active, with mixture-of-experts in every layer. It is natively multimodal: vision is trained in from the start with no separate image encoder. The training mix was 50T tokens with a March 2026 knowledge cutoff, and image-and-text data made up 18% of it. Microsoft says every token came from its own pipelines and none from distilling other models' outputs.

HOW IT WAS TRAINED

On 16,000 Nvidia GB200 GPUs over 38 days in August and September 2026, with an average 21% model FLOPs utilization and 89% goodput, by Microsoft's figures. It used the Muon optimizer rather than AdamW, and Microsoft's unique training data grew about 36% from MAI-Base-1, its June base model.

MICROSOFT'S OWN COMPARISON

Microsoft compares it only with open-weight models: DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. On its held-out code, STEM, math and Q&A sets, measured in bits per byte, it says MAI-Base-2 beats all of them. It also says its recipe is 4.2 to 12.3 times more compute-efficient than DeepSeek's. Three of those four are post-trained models, which Microsoft notes can skew a base-model test. These are Microsoft's numbers, and no closed model is in the comparison.

WHERE IT GOES

Into Microsoft's own products first, Microsoft says, then to Foundry customers, with the weights licensed to trusted partners. Microsoft's current flagship is MAI-Thinking-1, from June, post-trained on MAI-Base-1. Work on MAI-Base-3 has started.

WHAT IT MEANS FOR THE RADAR

Our MAI-2 hunt only took a bare mai-2 id, so a mai-base-2 or mai-thinking-2 listing would have slipped past. We widened it today to any MAI name ending in 2, outside the Image and Voice lines, and set it to alert our main list.

WHAT WE DO NOT KNOW

The name of the post-trained model, when it ships, its price, and how it scores against GPT-6 Astra, Claude Opus 5.5 or Gemini 4 Argon. Until it is post-trained and independently tested, nothing here says where it lands among the frontier models.