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

Microsoft AI said on October 15, 2026 that it has completed pre-training MAI-Base-2, a mixture of experts foundation model with 2.8 trillion total parameters and 80 billion active per token. Secondary summaries of the announcement report training on 16,000 GB200 GPUs over 38 days. The model is still in post-training, and Microsoft says it plans to license the weights to partners through Foundry, with no date or terms announced.

Sourcing note: this report rests on a single reporting chain, Microsoft's own Lab Notes post and summaries of it. The performance claims in that post, including that the model beats every open-weight model Microsoft tested and is 4 to 12 times more compute efficient than DeepSeek-V4-Pro-Base, are Microsoft's own estimates and have not been independently replicated. This item will be updated when independent evaluation or a release appears.

Sources: Microsoft AI; Vandata Team summary