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

Microsoft AI has finished pre-training MAI-Base-2, its next in-house foundation model. It is a mixture-of-experts model with about 2.8 trillion total parameters and 80 billion active per token, which puts it among the largest base models any company has disclosed.

Microsoft's main claim is efficiency. It says MAI-Base-2 gets 4 to 12 times more out of its pre-training compute than DeepSeek-V4-Pro-Base, and that it matches that model on held-out code with about 12 times less training compute. Those are Microsoft's own numbers, and outside evaluations are not out yet.

This is a base model, not a product. In June Microsoft announced MAI-Thinking-1, a reasoning model post-trained on top of MAI-Base-1, and the expectation is that something similar will follow here. Post-training results have not been published yet. Every step Microsoft takes on its own models also reduces how much it depends on outside labs for the AI inside its products.

Our take

The base model is the foundation, not the house. What matters is what Microsoft ships on top of it, and whether it starts replacing partner models inside Copilot and Azure.