Microsoft AI says it has pretrained MAI Base 2, a 2.8 trillion parameter base model built entirely in house
We now have state-of-the-art pre-trained base models in the 3T-parameter weight class, built entirely in-house at MAI. Post-training results coming soon -- stay tuned! Lots of details in our technical blog post.
Context
Microsoft building its own frontier scale model, rather than relying only on OpenAI, changes its position in the AI race and gives it models it can license through Foundry.
This is a first hand post from Mustafa Suleyman, CEO of Microsoft AI, linking a technical blog. The base model is not yet post trained or released, and the comparisons with open weight models are Microsoft's own evaluations.
Deep dive
On October 15, Mustafa Suleyman posted that Microsoft AI now has state of the art pretrained base models in the 3 trillion parameter weight class, built entirely in house at MAI, with post training results coming soon.
The linked blog names the model MAI Base 2, a mixture of experts model with 80 billion active and 2.8 trillion total parameters. Microsoft says it trained it on 16,000 GB200 GPUs over 38 days in August and September 2026, on a 50 trillion token mixture with a March 2026 knowledge cutoff, and that it was trained natively with vision.
Microsoft says MAI Base 2 scores better than every open weight model it tested on held out evaluations, including DeepSeek V4 Pro Base, Kimi K3 and Qwen3.8, and that its recipe is 4 to 12 times more efficient in training compute than DeepSeek's. It notes that only post trained versions of most rival models were available, which may affect the comparison.
The blog says Microsoft's pretraining uses organic data without distillation from other models, which it says will let it license model weights to trusted partners through Foundry. It says preparations for MAI Base 3 are already underway.
What to expect next
Watch for the post trained version of MAI Base 2 and how it scores on public benchmarks against models from OpenAI, Google and Anthropic.
Also watch which partners get licensed weights through Foundry, how Microsoft uses its own models alongside OpenAI's in Copilot, and when MAI Base 3 begins training. Independent tests of the model once it is available will show whether Microsoft's efficiency claims hold up.