The training run is the story—for now
Microsoft AI says it has finished pretraining MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active. The run used 16,000 GB200 GPUs for 38 days. The model is still undergoing post-training, so Microsoft has not yet presented it as a finished, generally available assistant. The current release is a technical account of how the lab trained it and what it measured.
Microsoft reports that its training recipe was 4.2 to 12.3 times more FLOPs-efficient than the reference open base model across four held-out evaluations. That is the company’s estimate of the compute needed for its recipe to reach a comparable loss—not a claim that a user-facing MAI model is 12 times faster or better. Microsoft says the efficiency came from scaling methods, model-infrastructure co-design, and changes to mixture-of-experts routing and optimization.
What builders should watch next
The lab says MAI-Base-2 was pretrained on 50 trillion tokens, including 4.5 trillion unique vision-language tokens. Its training data had a March 2026 knowledge cutoff, and the company says it is built from data processed in-house rather than distilled from other models. Those details matter because they will shape what the model can do once post-training and partner access arrive.
For now, this is evidence of a scaled training pipeline, not yet a public capability demo. The next useful comparison will be the post-trained model’s published evaluations and real application behavior. Microsoft says it plans to license weights to trusted partners through Foundry after post-training; availability and independent testing remain the milestones to watch.
Source published 2026-10-15. Coverage is based on the maker’s announcement and demonstration.