Microsoft finishes pre-training MAI-Base-2 foundation model
Microsoft AI said Thursday it finished pre-training MAI-Base-2, a mixture-of-experts model with 80 billion active and 2.8 trillion total parameters, trained on 16,000 Nvidia GB200 GPUs. The model is in post-training; Microsoft says it may later license weights to partners through Foundry.
AI-assisted · US Brief desk · Sources listed below
What happened
Microsoft AI said Thursday it has finished pre-training MAI-Base-2, describing it as a mixture-of-experts foundation model with 80 billion active parameters and about 2.8 trillion total parameters. The company said the run used 16,000 Nvidia GB200 GPUs over 38 days in August and September 2026, with average model FLOPs utilization of 21% and goodput of 89%.
Why it matters
The announcement advances Microsoft's push, led by Microsoft AI CEO Mustafa Suleyman, to own more of its model stack after years of relying heavily on OpenAI and Nvidia. Microsoft said finishing the base run lets it license weights to trusted partners through Foundry once post-training is done, and to tune models with its own Azure and Maia infrastructure.
More context
What Microsoft claims: On its held-out evaluations, Microsoft said MAI-Base-2 beat every open-weight model it tested and was about 4–12× more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, the strongest open base model in its comparison set. It also said the training used organic data without distillation from other models, and that the model was pre-trained natively with vision.
Primary excerpt: "We've recently completed pre-training our new state-of-the-art model, MAI-Base-2, an MoE foundation model with 80 billion active and 2.8 trillion total parameters," Microsoft AI wrote. "Owning our model pipeline end-to-end allows Microsoft to make a distinctive offering to our customers… This will enable us to license our model weights to trusted partners through Foundry."
Short timeline and what's still unknown: Microsoft introduced MAI-Base-1 and MAI-Thinking-1 around its June 2026 Build lineup; MAI-Base-2 trained in August–September; pre-training results posted Oct. 15. The company said MAI-Base-2 is now in post-training and that work on MAI-Base-3 has started. It has not published public weights, a consumer release date, or third-party confirmation of the benchmark claims.
3 listed sources
- Microsoft AI, Oct. 15, 2026: Frontier Pre-training at Microsoft AI
- Microsoft AI PDF, Oct. 15, 2026: Frontier Pre-training at Microsoft AI
- Crypto Briefing, Oct. 15, 2026: Microsoft finishes pre-training MAI-Base-2
References listed by US Brief; a source count is not a verification score.
Editorial sourcing notes
Training scale, efficiency, and evaluation claims are from Microsoft AI's Oct. 15, 2026 technical post and PDF. They are Microsoft's own held-out comparisons, not an independent third-party benchmark. Crypto Briefing republished the announcement the same day.
Like US Brief? You can support it with a tip.