MAI-Base-2
Microsoft AI's 2.8T-parameter MoE base model, pre-trained on 16K GB200 GPUs in 38 days; in post-training, not yet released.
Overview
Microsoft AI announced on October 15, 2026 that it finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total and 80 billion active parameters. It was trained on 16,000 GB200 GPUs over 38 days on a 50T-token mixture, with 21% average MFU and 89% goodput. Microsoft says it beats every open-weight model it tested on held-out evaluations and is 4.2 to 12.3 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, based on its own estimates. The model is still in post-training and not available to customers; Microsoft plans to license weights to trusted partners through Foundry, and work on MAI-Base-3 has started.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- 2.8T total / 80B active parameters, mixture-of-experts
- Trained on 16K GB200 GPUs over 38 days; 50T-token mixture with 18% vision-language data
- Microsoft-reported 4.2x to 12.3x pre-training FLOPs efficiency versus DeepSeek-V4-Pro-Base
- Built on organic data with no distillation from other models, per Microsoft
- Planned weight licensing to trusted partners through Foundry
Use cases
- Base for Microsoft's own post-trained MAI models
- Licensed weights for trusted partners via Foundry (planned)
Pricing
unknown; not released to customers.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 15 Oct 2026
- Record updated
- 18 Oct 2026
- Record last checked
- 18 Oct 2026