Microsoft AI said it has finished pre-training MAI-Base-2, a mixture-of-experts model, which switches on only part of its network for each input. It has 2.8 trillion parameters in total and 80 billion active at once. Microsoft said the run lasted 38 days, from August into September 2026, on 16,000 GB200 chips. It said the model beat every open-weight model it tested on bits-per-byte, a score where lower is better, across four held-out evaluations. Microsoft ran the tests itself.
Microsoft estimates its training recipe is 4.2 to 12.3 times more compute-efficient than DeepSeek-V4-Pro-Base. That is an estimate, not a measurement. Microsoft fitted scaling laws on smaller models, then worked out how much compute its recipe needs to match DeepSeek's score. MAI-Base-2 itself used more compute than DeepSeek's model. Microsoft credits choices such as mixture-of-experts layers in every layer and a memory fix for overloaded experts. The announcement gives no independent verification. It also gives no post-training results, release date, pricing or partner list.
This is a claim about training efficiency, not about a product you can use today. The weights are not open: Microsoft plans to license them to trusted partners through Foundry. Buyers should ask what post-training results show, what licence terms and pricing apply, and who gets access.