Microsoft trains giant model MAI-Base-2: 2.8 trillion parameters in 38 days
Microsoft's AI division has completed pre-training of its largest in-house model to date. It is said to compute far more efficiently than DeepSeek – but will not be released openly.
Archive image · Photo: Christopher Wilson / Wikimedia Commons, CC BY-SA 4.0 (cropped and resized)
Microsoft AI has completed pre-training of MAI-Base-2, the group's AI division announced on Thursday. The mixture-of-experts model has 2.8 trillion parameters, 80 billion of them active. It was trained for 38 days in August and September on 16,000 Nvidia GB200 chips using 50 trillion tokens, 18 per cent of them image-text data. Its knowledge runs to March 2026.
By Microsoft's own calculations, MAI-Base-2 is 4.2 to 12.3 times as compute-efficient as the DeepSeek-V4-Pro base model across four tests. Microsoft could only measure other comparison models in their post-trained versions, which may skew the results. According to Microsoft, the model was not created by distilling other companies' models.
Post-training is now under way. Microsoft does not plan to release the weights openly but to license them to trusted partners via its Foundry platform; the company named no partners or dates. The predecessor MAI-Base-1, with one trillion parameters, was the basis for MAI-Thinking-1 in June. Work on MAI-Base-3 has already begun.