The company said it finished pretraining MAI-Base-2, a 2.8-trillion-parameter model trained on 16,000 Nvidia GB200 GPUs. (Oct. 17)

Microsoft AI said on Oct. 15 that it had completed pretraining MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active per token. The August–September run used 16,000 Nvidia GB200 GPUs for 38 days and processed 50 trillion tokens; Microsoft reports average model-FLOPs utilization of 21% and training goodput of 89%. The company says the model beat every open-weight system it tested on held-out evaluations and its pretraining recipe was 4.2–12.3 times more FLOPs-efficient than DeepSeek-V4-Pro-Base across four tests. Those are company-reported base-model results, not independent evidence of finished assistant performance: post-training is still underway, and Microsoft has not announced pricing or a release date. Unlike MAI-Base-1, announced in June with one trillion parameters, MAI-Base-2 includes vision in pretraining, with image-related data making up 18% of its mixture.