Microsoft Finishes Pre-Training MAI-Base-2, a 2.8-Trillion-Parameter In-House Model
Microsoft AI says it has finished pre-training MAI-Base-2, a mixture-of-experts model with 80 billion active and 2.8 trillion total parameters, and claims it beats every open-weight model it tested. Post-training is underway; no release date yet.
By 1EyeAI Desk ·
Microsoft AI said on Oct. 15, 2026 that it has completed pre-training MAI-Base-2, a mixture-of-experts foundation model with 80 billion active and 2.8 trillion total parameters. It follows MAI-Base-1, the base for the MAI-Thinking-1 reasoning model announced in June.
Microsoft says it trained the model on 16,000 NVIDIA GB200 GPUs over 38 days in August and September 2026, on a 50-trillion-token mix with a March 2026 knowledge cutoff. Unlike its predecessor, it was trained natively with images, which made up 18% of the mix. Microsoft says the data is organic rather than distilled from other models, and that this will let it license the weights to trusted partners through Foundry. That is a stated plan.
Microsoft claims MAI-Base-2 beat every open-weight model it tested on its own held-out code, STEM, math and Q&A evaluations, including DeepSeek-V4-Pro-Base, Kimi K3 and Qwen3.8, and that its recipe is 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base’s. These are company claims on Microsoft’s own tests, not independent benchmarks, and several rivals were compared using post-trained rather than base checkpoints.
The model is now in post-training, and Microsoft says work on MAI-Base-3 has begun. It has not said when a finished model will ship.
What this means for you: Nothing to use yet. If you build on Azure or Microsoft Foundry, expect new first-party Microsoft models built on this base, and test them on your own tasks against the models you already use once they arrive.