Microsoft finishes pre-training MAI-Base-2, a 2.8T-parameter AI model with 4-12x better compute efficiency

Microsoft has completed pre-training for its new foundation model called MAI-Base-2. The model has 2.8 trillion total parameters, with 80 billion active ones. The company says it beats all open-weight models tested in performance.

The model was trained on 6,000 NVIDIA GB200 GPUs over 38 days. Its average compute utilization was 21%, with 89% effective throughput. Microsoft says its training efficiency is 4 to 12 times better than DeepSeek-V4-Pro-Base.

Microsoft credits three key moves: scientific scaling, model-infrastructure co-design, and data scaling. The company has already started preparing its next model, MAI-Base-3.