Microsoft finishes training MAI-Base-2, a 2.8-trillion-parameter model it says needs up to 12 times less compute than DeepSeek’s

What Microsoft announced

Microsoft AI, the division led by Mustafa Suleyman, said on 15 October 2026 that it has completed pre-training of MAI-Base-2, its next in-house foundation model.¹² It is a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active.¹²

How it was trained

According to Microsoft:¹³

• Training ran for 38 days in August and September 2026 on 16,000 Nvidia GB200 GPUs.
• It used a 50-trillion-token training mix, with images making up 18%, so vision is built in from the start.
• The model was trained on original data rather than by distilling other companies’ models.

What Microsoft claims

Microsoft says MAI-Base-2 is 4.2 to 12.3 times more compute-efficient in pre-training than DeepSeek-V4-Pro-Base across four evaluations, and outperforms every open-weight model it tested.¹² It says the training recipe is 6.3 to 12.7 times more efficient than its predecessor, MAI-Base-1.¹ These are Microsoft’s own comparisons, made against base and post-trained checkpoints that it notes may not be like for like.¹

When and how it will be offered

MAI-Base-2 is still in post-training, and Microsoft gave no release date.¹² It plans to license the model weights to trusted partners through Foundry.¹² GIGAZINE notes the earlier MAI-Base-1 was closed-weight.³ Microsoft says work on MAI-Base-3 has begun.¹

Why it matters

Microsoft is building its own frontier models alongside its partnership with OpenAI. It is also building its own chips; see our story on Maia 300.