Microsoft has completed pre-training of MAI-Base-2, a mixture-of-experts (MoE) foundation model with 80 billion active parameters and 2.8 trillion total parameters, according to a statement from Microsoft AI.

The model was trained on 16,000 GB200 GPUs over 38 days in August and September 2026. Microsoft says MAI-Base-2 outperforms every open-weight model it tested on held-out evaluations and is 4 to 12 times more pre-training FLOPs-efficient than DeepSeek-V4-Pro-Base, which the company describes as the strongest available open base model.

MAI-Base-2 follows MAI-Base-1, which Microsoft announced in June 2026, and on which it built MAI-Thinking-1, a general-purpose reasoning model. The new model's training corpus grew by roughly 36% compared to MAI-Base-1, measured in unique training tokens, and includes a 50 trillion-token training mixture with a March 2026 knowledge cutoff.

Unlike its predecessor, MAI-Base-2 is pre-trained natively with vision using an encoder-free architecture. Its vision-language model corpus contains 4.5 trillion unique tokens, accounting for 18% of the total training mixture. Tasks covered include image captioning, chart understanding, spatial reasoning, and medical understanding.

Microsoft said the model's training data is built on organic data and does not involve distillation from other models. The company said this approach will allow it to license model weights to enterprise customers through its Foundry platform.

MAI-Base-2 is currently undergoing post-training. Microsoft said preparations for MAI-Base-3 are already underway.