Microsoft completes pretraining of MAI-Base-2

Microsoft AI says it has completed pretraining MAI-Base-2, a mixture-of-experts foundation model with 2.8 trillion total parameters and 80 billion active parameters per token. This is a training milestone, not a user-facing launch. Microsoft says the model is now undergoing post-training and that it will share more results later. The company plans to license the weights to trusted partners through Foundry, but it has not announced a public download, API price or release date.

The training run shows the scale of Microsoft’s attempt to build more of its own AI stack. The company says it used 16,000 GB200 GPUs for 38 days in August and September 2026. The model was trained on a 50-trillion-token mixture with a March 2026 knowledge cutoff. Vision was included during pretraining rather than added later: Microsoft reports 4.5 trillion unique visual tokens, which represented 18 percent of the mixture when repeated data was counted. These are Microsoft’s descriptions of its own infrastructure and data pipeline.

Microsoft compares MAI-Base-2 with DeepSeek-V4-Pro-Base, Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL on held-out pretraining evaluations. It says MAI-Base-2 achieved lower bits-per-byte scores and estimates that its recipe was 4.2 to 12.3 times more training-FLOPs-efficient than DeepSeek’s base model. Those results are not an independent ranking of finished assistants: Microsoft selected the tests, the comparison includes post-trained checkpoints and the model itself is not yet post-trained. Independent coverage from ArtificialWatch and CellCog therefore treats the efficiency figures as Microsoft claims awaiting outside replication.

The announcement matters because Microsoft is trying to own more of the path from data and model training to chips, products and enterprise distribution. A successful post-trained model could reduce reliance on outside providers and give Microsoft another option for Copilot and Foundry customers. It could also change the economics of long-running agents if the training recipe translates into efficient inference, but that has not been demonstrated yet.

For users and makers, the practical news is what to watch next: access, independent evaluations, safety testing and real-world costs. A very large base model is not automatically a useful assistant. The milestone is significant for the market, while its value for everyday AI work will depend on what Microsoft releases after post-training.