Microsoft AI details MAI-Base-2, a 2.8-trillion-parameter model trained in-house

Microsoft AI published a technical note saying it had finished pre-training MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters, 80 billion of them active for any token. It is the successor to MAI-Base-1, on which the company built its June reasoning model MAI-Thinking-1. The model is still in post-training, and Microsoft says it will share more later.

The note is a pre-training report, not a product launch, so its claims concern base-model quality. Microsoft says MAI-Base-2 “outperforms every open-weight model we tested” on held-out evaluations measured in bits per byte, a tokenizer-independent loss in which lower is better. The test sets cover code from Microsoft’s own codebase and STEM, maths and question-answering sets commissioned from vendors, decontaminated against its training data. The comparison set was open-weight models: DeepSeek-V4-Pro-Base, the only open base checkpoint above a trillion parameters, plus Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL, for which only post-trained checkpoints exist, a caveat Microsoft flags. No proprietary frontier model is in the comparison, and no post-trained benchmark results are given.

Microsoft also says its recipe is 4.2 to 12.3 times more efficient in pre-training compute than DeepSeek-V4-Pro-Base, estimated by fitting scaling laws to its own smaller models, and that it spent the saving on a larger model. It reports training on 16k GB200 GPUs for 38 days in August and September 2026, at 21% model-FLOPs utilisation and 89% goodput, on a 50-trillion-token mixture. Unique training tokens grew by roughly 36% over MAI-Base-1, with vision data added natively. The engineering described includes the Muon optimiser, a dropless mixture-of-experts layer and clipping of activation gradients.

The commercial argument is data provenance. Microsoft says the model is built on organic data without distillation from other models, and that this will let it license weights to “trusted partners” through its Foundry platform. The company frames the work as owning the pipeline end to end, and says preparations for MAI-Base-3 are under way.

The note is Microsoft’s own, the evaluations are internal, and no weights or post-trained results were released with it. The run used Nvidia GB200 GPUs; Microsoft had unveiled its third Maia chip three days earlier. What the note establishes is the scale at which Microsoft’s in-house lab is now pre-training, and its stated plan to license the weights to partners as well as use them in its own products.