Microsoft finishes pre-training MAI-Base-2, a 2.8 trillion parameter model

Microsoft announced it has finished pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active for each input. It follows MAI-Base-1, introduced earlier this year.

Microsoft says it trained the model on 16,000 Nvidia GB200 GPUs over 38 days in August and September, on 50 trillion tokens of text and images. Images made up 18% of the training mix. It estimates its approach was 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base at matching performance on selected tests, and says the model learned from organic data rather than by distilling other AI models.

The caveats are real. The comparisons use selected internal evaluations, and Microsoft has not published equivalent results against every leading proprietary model. The model is still in post-training, with no pricing or release date.

Microsoft plans to license the weights to trusted partners through its Foundry platform rather than release them publicly. That is the opposite of an open-weight release.

Our take

Microsoft is building its own frontier model while its relationship with OpenAI evolves. The licensed-weights approach is a middle path between closed APIs and open downloads.