Microsoft said it had completed pre-training its 2.8-trillion-parameter MAI-Base-2 model, claiming it outperformed the open-weight models it tested while requiring less computing power.
Microsoft AI said on Oct. 15, 2026, that it had completed pre-training MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active per token. The company trained it on 16,000 Nvidia GB200 GPUs for 38 days in August and September, using about 50 trillion tokens; the model is now undergoing post-training. Microsoft says MAI-Base-2 beat every open-weight model it tested on held-out evaluations in code, question-answering, math and STEM. Its scaling-law estimates put the training recipe at 4.2–12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base at comparable evaluation scores—not a claim that the full run used less compute. The results are company-reported, and Microsoft has not announced public weights or pricing. The milestone strengthens Microsoft’s bid to build frontier models in-house, potentially giving its Copilot and cloud businesses more control over AI supply as it competes with OpenAI and open-model developers.