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

On October 15, Microsoft AI said it has finished pre-training MAI-Base-2, an in-house mixture-of-experts model with 80 billion active parameters and 2.8 trillion total. Microsoft says it trained the run on 16k GB200 GPUs over 38 days in August and September 2026. Mustafa Suleyman posted that post-training results are "coming soon."

These next lines are Microsoft's claims, from that post, not an independent bench. Microsoft says MAI-Base-2 outperforms every open-weight model it tested on held-out evaluations, scored as bits per byte, lower better. It says the recipe is 4.2 to 12.3 times more pre-training-FLOPs-efficient than DeepSeek-V4-Pro-Base. The opening summary rounds that to 4–12 times. Microsoft also says the pre-training uses organic data and does not distill from other models, and that it plans to license the weights to trusted partners through Foundry. Post-training is underway. There is no public weight, no API, and no price. StreetInsider carried the completion.

What it costs

Nothing to buy. Microsoft has not published a price for MAI-Base-2.

Should you switch

No. There is nothing to call yet. It matters because Microsoft is training its own frontier base model rather than only serving OpenAI's. If a paid version ships, it lands on the prices page only if it clears the inclusion rule already printed there: within about 10 points of the Artificial Analysis Intelligence Index leader, or in the top 15 on LMArena text, and a published paid price.