Microsoft AI’s Oct. 15 technical update puts MAI-Base-2 at 2.8 trillion total parameters, with 80 billion active, and marks the completion of its pretraining. The company reports training the mixture-of-experts model on 16,000 GB200 GPUs for 38 days during August and September 2026.
Microsoft’s comparisons use held-out text in code, question-answering, math and STEM, measured in bits per byte, where lower scores are better. It reports beating every open-weight model it tested on those evaluations. The comparison includes post-trained versions of Kimi K3, Qwen3.8-2.4T-A95B and MiMo-V2.6-Pro-RL, and Microsoft cautions that post-training can affect the results.
The company’s claimed 4–12 times improvement in pretraining compute efficiency has a specific scope. Microsoft fitted scaling laws to estimate how much compute its recipe would need to match DeepSeek-V4-Pro-Base’s scores, yielding a range of 4.2–12.3 times across the four evaluations. That is a modeled efficiency comparison, not a claim that the full MAI-Base-2 run used less total compute: Microsoft notes that DeepSeek’s base model was trained with significantly fewer floating-point operations.
These are Microsoft’s own pretraining results. MAI-Base-2 is undergoing post-training, the company says, with further results still to come.
