MAI-Base-2: Microsoft Finishes a 2.8T-Parameter Model

Microsoft AI has finished pre-training MAI-Base-2, the base model it will build its next generation of in-house AI on. In a technical post published on 15 October, the team describes a mixture-of-experts model with 2.8 trillion parameters in total and 80 billion active per token, trained on 16,000 Nvidia GB200 GPUs over 38 days in August and September. It is now in post-training, the stage that turns a raw base model into an assistant, and Microsoft has not said when, or under what name, the finished version will ship.

The jump from its predecessor is large. MAI-Base-1, the base under the MAI-Thinking-1 reasoning model Microsoft released in June, has about a trillion parameters with 35 billion active, according to Microsoft’s own model page. The new model puts experts in every layer rather than every other one, and it is natively multimodal: images are trained in from the start through an encoder-free design, with vision-language data making up 18% of a 50-trillion-token training mix. Its knowledge stops in March 2026.

The efficiency claim

The number Microsoft leads with is efficiency. It says its pre-training recipe is 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base, which it calls the strongest open base model and the only released base checkpoint above a trillion parameters. The method is worth spelling out, because it is not a head-to-head score. For each of four held-out test sets, Microsoft fitted a scaling law of loss against training compute from its own ladder of smaller runs, then estimated how much compute its recipe would need to match DeepSeek’s result and compared that with what DeepSeek actually spent. Against MAI-Base-1 it claims a 6.3 to 12.7 times gain since June, and it says it spent the savings on training a bigger model.

The comparisons use bits per byte, a measure of how well a model predicts text that does not depend on its tokenizer, on four held-out sets: code from Microsoft’s own codebase, plus STEM, maths and question-answering sets commissioned from vendors and checked against the training data for contamination. On all four, Microsoft says, MAI-Base-2 beats every open model it tested with more than a trillion parameters: DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL. Only DeepSeek’s model is a true base checkpoint; the other three are post-trained releases, which Microsoft concedes may skew the result. These are Microsoft’s own internal tests, and the post gives its results only as charts.

How it was built

The post is unusually specific about the training run. Average model FLOPs utilisation was 21% and goodput 89%, and work on the expert layers lifted their utilisation from 24% to 35%. The team used the Muon optimiser and published the scaling laws it uses to set learning rate, weight decay and dropout. On data, Microsoft says every token is “organic”, with no distillation from other models, that its own crawler respects robots.txt, and that it excludes sites on the US Trade Representative’s Notorious Markets list of piracy and counterfeiting hubs. The unique-token corpus grew by about 36% over MAI-Base-1.

What Microsoft will do with it is sketched in one line: it wants to license the weights to trusted partners through Foundry, tuning models for its own products first and Foundry customers second. That is not an open-weight release, and the post names no partners, prices or dates; GIGAZINE, reporting on the post, noted that it leaves the model’s openness unstated. It landed the same day Meta put out Muse Spark Open under its own license, and it shows the other route a US lab can take: own the base, keep the weights, and rent them out. Preparations for MAI-Base-3, the post adds, are already under way.