Microsoft completes pre-training MAI-Base-2

Microsoft describes MAI-Base-2 as a new mixture-of-experts foundation model with 80 billion active parameters and 2.8 trillion total parameters.[1] The company says it is larger, sparser and trained on more data than MAI-Base-1.[1] Microsoft also says it pretrained MAI-Base-2 natively with vision.[2] TokenPost reports that the training mixture contained 50 trillion tokens, of which vision-language data accounted for 18%, or 9 trillion tokens.[2] Microsoft has not announced partners or a release date.[2]

Microsoft says the training run used 16,000 GB200 GPUs over 38 days in August and September 2026.[1] It reports average model FLOPs utilization of 21% and goodput of 89%.[1] Microsoft says its training recipe was 4.2–12.3 times more FLOPs-efficient, and that progress since the model’s release led to a 6.3–12.7 times FLOPs-efficiency improvement.[1] Those are Microsoft’s estimates, not independent measurements established by the evidence here. Microsoft says it decontaminated all evaluation sets against its training data.[1]

Microsoft says MAI-Base-2 outperformed every open-weight model it tested on held-out evaluations.[1] It compared the model with frontier checkpoints available before October 2026 and identifies DeepSeek-V4-Pro-Base as the most capable open-weight base model in its comparison.[1] The company cautions that some models were available only as post-trained checkpoints, and that post-training may affect evaluation results.[1] The reported results therefore do not settle how all models would compare under identical conditions.

Microsoft says MAI-Base-2 is undergoing post-training and that it expects to share more results later.[1] It plans to license the model’s weights to trusted partners through Foundry.[1] TokenPost reports that Microsoft has not announced any partners or a release date.[2] TestingCatalog separately reports that preparation for a MAI-Base-3 training cycle is already underway.[3] For teams considering adoption, the immediate distinction is between a completed pre-training run and a model that is ready for external use: the evidence confirms the former, but gives no launch timetable or named partner.