Microsoft’s 2.8T-parameter model cuts pretraining compute 4 to 12x
Microsoft says MAI-Base-2 is 4 to 12 times more pre-training FLOPs-efficient than the strongest open base model it tested, giving Foundry customers a potentially cheaper route to licensed model weights. The mixture-of-experts model has 2.8 trillion total parameters but activates 80 billion at a time, and Microsoft trained it on 16k GB200 GPUs over 38 days. Post-training is still underway, so product-level results are not available yet.