Microsoft finishes pretraining MAI-Base-2, its own frontier base model
Microsoft AI has finished pretraining MAI-Base-2, a 2.8-trillion-parameter base model trained on 16,000 GB200s over 38 days. The model is now in post-training and will reach Microsoft products first, then Foundry customers.
In a lab note published on October 15, 2026, Microsoft AI announced that it has finished pretraining MAI-Base-2, a base model it built in-house. The company says the model is now in post-training, that it will share results soon, and that preparation has started on the next model, MAI-Base-3.
What MAI-Base-2 is
MAI-Base-2 is a mixture-of-experts base model with 2.8 trillion parameters in total, 80 billion of which are active at each step. A base model is the raw model before post-training for chat and instruction following; the assistant models used in products are built on top of it. MAI-Thinking-1, which Microsoft announced in June, was built on the previous version, MAI-Base-1. The new model learns images and text together from the start and uses no separate vision encoder. Its knowledge cutoff is March 2026.
How it was trained
The model was trained on 16,000 NVIDIA GB200 chips over 38 days in August and September 2026. The training mixture was 50 trillion tokens, about a fifth of it vision-language data. Microsoft says all data was processed in-house, that its web crawler respects robots.txt, and that the model was not trained on data distilled from other models. Sources on the US Trade Representative's Notorious Markets list were excluded.
Microsoft products first
Microsoft says the model will be used in its own products first and then offered to Foundry customers. It also plans to license the model files to trusted partners through Foundry. The note gives no indication that the model will be released openly.
Microsoft says MAI-Base-2 beat every open model it tested on its held-out evaluations and was trained far more efficiently for the same compute than DeepSeek-V4-Pro-Base, which it sees as the strongest open base model. The company itself notes that these evaluations are its own, that some were commissioned from vendors to reduce contamination, and that some of the compared models could only be measured in their post-trained versions. Independent test results will be worth watching once the model reaches products.
