Microsoft completes pre-training of MAI-Base-2: 2.8 trillion parameters, weights not yet released
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Microsoft showed that it trained its own base model at the 2.8-trillion-parameter scale, but for now it is unavailable.
According to Microsoft's description, MAI-Base-2 is a mixture-of-experts model with 2.8 trillion parameters, 80 billion of them active, with images handled in a built-in architecture without a separate encoder. Training on 50 trillion tokens took 38 days on 16,000 GB200 GPUs (August-September 2026). The company claims the model is 4-12 times more efficient in pre-training cost than the strongest open base model, DeepSeek-V4-Pro-Base, and beats the open models tested on held-out evaluations — these are Microsoft's own figures. Post-training is under way; the weights have not been published, and licensing is planned through Foundry.
Sources 1