Microsoft completes pre-training of 2.8-trillion-parameter MAI-Base-2

Microsoft AI has completed pre-training of MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active parameters, the company said. That milestone applies to the base model; MAI-Base-2 is still undergoing post-training before Microsoft shares further results.

The mixture-of-experts design uses 80 billion active parameters out of the larger 2.8-trillion-parameter model. Microsoft said it trained MAI-Base-2 on 16,000 GB200 GPUs for 38 days during August and September 2026. The company reported average model FLOPs utilization of 21% and goodput—the share of time in which training progressed—of 89%. Those operational figures have not been independently audited.

Microsoft said the training mixture totaled 50 trillion tokens. Vision-language data made up 18% of that mixture after repetition, while the underlying vision-language corpus contained 4.5 trillion unique tokens.

In its held-out evaluations, Microsoft said MAI-Base-2 recorded lower bits per byte than every open-weight model it tested. The company also estimated that its training recipe was 4.2 to 12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base. Both comparisons come from Microsoft’s own tests and scaling-law fits, not an independent benchmark. Microsoft also cautioned that several comparison models were available only as post-trained checkpoints, limiting direct comparison with a base model.

Microsoft said MAI-Base-2 will serve as a foundation for future models and that its data provenance will enable it to license the model weights to trusted partners through Foundry. The announcement did not provide a weight-availability date, pricing or detailed licensing terms.