Microsoft trains a 2.8-trillion-parameter model of its own and will share it only with partners

The news

Microsoft AI said it completed pre-training of MAI-Base-2, a mixture-of-experts model with 2.8 trillion total and 80 billion active parameters, trained on 16,000 GB200 GPUs over 38 days in August and September on 50 trillion tokens. The model is now in post-training, and Microsoft plans to license the weights to trusted partners through Foundry.

Why it matters

Microsoft has depended on OpenAI for its frontier models. A base model of this size gives it leverage, and sharing weights selectively, not openly, sets it apart from Meta and Nvidia.

The details

Microsoft claims MAI-Base-2 beats every open-weight model it tested on held-out bits-per-byte, a measure of how well a model predicts unseen text, and that it is 4.2 to 12.3 times more compute-efficient in pre-training than DeepSeek-V4-Pro-Base. The efficiency numbers are estimates from fitted scaling laws and are vendor-reported.

About 18 percent of the training data was vision-language material. The knowledge cutoff is March 2026.

What we could not check

• Base-model comparisons cannot be checked until partners get access.
• Microsoft compared some rivals only as post-trained checkpoints.