Microsoft AI finishes pre-training MAI-Base-2, a 2.8-trillion-parameter model trained on 16,000 GB200 chips in 38 days

Microsoft AI said on 15 October that it has finished pre-training MAI-Base-2, “an MoE foundation model with 80 billion active and 2.8 trillion total parameters”. “We trained MAI-Base-2 on 16k GB200 GPUs over 38 days in August and September 2026,” the lab's post says. It was trained on a mixture of 50 trillion tokens, 18% of it vision-language data, with no distillation from other models.

The lab claims its recipe is 4.2 to 12.3 times more efficient in pre-training compute than DeepSeek-V4-Pro-Base, which it calls the strongest open base model. It also says MAI-Base-2 beats every open-weight model it tested on held-out evaluations. Both claims rest on the lab's own measurements, including bits-per-byte comparisons and fitted scaling laws, and have not been checked independently.

The model is still in post-training and has no release date. Microsoft plans to license the weights to trusted partners through Foundry, its platform for AI developers, and says work on MAI-Base-3 has begun. Reports that the lab's chief, Mustafa Suleyman, told staff the model beats OpenAI's base model are not on the post and are left out here.

The announcement came three days after Microsoft unveiled its Maia 300 inference chip.

• Confirmed MAI-Base-2 has 2.8 trillion total and 80 billion active parameters and was trained on 16k GB200 GPUs over 38 days. Microsoft AI
• Claimed Microsoft AI claims its recipe is 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base. Microsoft AI
• Confirmed The model is in post-training, and its weights are to be licensed to trusted partners through Foundry. Microsoft AI