MAI-Base-2
Microsoft AI announced MAI-Base-2, a 2.8 trillion parameter mixture-of-experts base model with 80 billion active parameters, which finished pre-training and is now in post-training.
MAI-Base-2 follows MAI-Base-1 from June. Its unique token dataset grew about 36%, and training now includes vision natively, with images making up 18% of the training mixture. Microsoft says its recipe is 6.3 to 12.7 times more compute-efficient than MAI-Base-1 and 4.2 to 12.3 times more efficient than DeepSeek-V4-Pro-Base across four evaluations.
- Date
- Thursday 15 October 2026
- Lab
- Microsoft
- Kind
- model
- Access
- paper only
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total parameters | 2.8 trillion mixture of experts | company |
| Active parameters | 80 billion per token | company |
| Training hardware | 16,000 Nvidia GB200 GPUs for 38 days August and September | company |
| Training tokens | about 50 trillion | company |
| Compute efficiency vs DeepSeek-V4-Pro-Base | 4.2 to 12.3 times across four evaluations | company |
| Compute efficiency vs MAI-Base-1 | 6.3 to 12.7 times | company |
Nobody can test it yet. Microsoft plans to license weights to trusted partners through Azure AI Foundry, with no dates or pricing given. Access is set to the closest allowed value.
Sources
- sg.finance.yahoo.com/news/microsoft-ai-unveils-2-8-165617565.html
- pasqualepillitteri.it/en/news/24354/microsoft-mai-base-2-trillion-parameters
This record was checked against its sources on 15 October 2026. How we check