Microsoft MAI-Base-2: 2.8T-Parameter Model Trained on 16K GB200s

Microsoft AI has finished pre-training MAI-Base-2, an in-house mixture-of-experts model with 2.8 trillion total and 80 billion active parameters. The run took 38 days on 16,000 NVIDIA GB200 GPUs.

Mustafa Suleyman announced the milestone on October 15, 2026, describing it as a base model in the “3T-parameter weight class” built entirely in-house. Microsoft’s technical post says the run took place in August and September. It averaged 21% MFU (model FLOPs utilization) and 89% goodput.

MAI-Base-2 is larger and sparser than MAI-Base-1, which had 1T total and 35B active parameters. TokenPost reports that MAI-Base-2 was trained on a 50-trillion-token mix, 18% of it vision-language data, and that it was pretrained natively with vision. Microsoft credits the gains to scaling-law work, Muon-based optimizer improvements and better data pipelines.

The model is now in post-training. Microsoft plans to license the weights to trusted partners through Foundry. It hasn’t announced partners or a release date. The result shows Microsoft building frontier capacity of its own alongside its partnership with OpenAI.