Microsoft Completes Pretraining of 2.8 Trillion-Parameter MAI-Base-2

Microsoft announced Oct. 15 that it had completed pretraining MAI-Base-2, a 2.8 trillion-parameter model, after a 38-day run on 16,000 NVIDIA GB200 GPUs, moving it into post-training.

The training campaign ran during August and September 2026. MAI-Base-2 has 80 billion active parameters, and its full parameter count is 2.8 trillion. Its knowledge cutoff is March 2026.

The training mixture contained 50 trillion tokens, with vision-language data accounting for 18%. That share represents 9 trillion tokens in the mixture; the vision-language corpus contained 4.5 trillion unique tokens. The model was pretrained natively with vision.

MAI-Base-2 is larger and sparser than MAI-Base-1, which had 35 billion active parameters and 1 trillion total. MAI-Base-1 served as the base model for MAI-Thinking-1, a reasoning model announced in June. Training for the earlier model used 8,192 NVIDIA GB200 GPUs and included 30 trillion pretraining tokens followed by 3.55 trillion mid-training tokens. The newer run used nearly twice as many GPUs.

The estimated FLOPs efficiency of the MAI-Base-2 training recipe was 4.2 to 12.3 times that of DeepSeek-V4-Pro-Base across four held-out evaluations. The broader comparison included Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL, for which only post-trained checkpoints were available. That checkpoint difference may affect those results.

MAI-Base-2 is the foundation for Microsoft's next generation of models, with preparations for MAI-Base-3 underway. Microsoft plans to license MAI-Base-2's weights to trusted partners through Foundry. No partners or release date have been announced, and the model remains in post-training.