Microsoft says it has completed pre-training MAI-Base-2, a new artificial intelligence model with 2.8 trillion total parameters. The model is still undergoing post-training, and Microsoft has not announced a public release date.
MAI-Base-2 reaches frontier scale
MAI-Base-2 uses a mixture-of-experts design, in which different parts of the model handle different tasks. It has 80 billion active parameters and 2.8 trillion total parameters, according to Microsoft’s announcement.
Microsoft says MAI-Base-2 performed better than every open-weight model included in its held-out evaluations. The company also says its training method was 4.2 to 12.3 times more efficient in computing operations than the method used for DeepSeek-V4-Pro-Base. These comparisons come from Microsoft’s own evaluations.
Training required major infrastructure
Microsoft trained the model on 16,000 GB200 GPUs over 38 days during August and September 2026. The training run used a 50-trillion-token mixture, while the unique training corpus grew by roughly 36% compared with MAI-Base-1.
Microsoft also trained MAI-Base-2 with vision from the beginning, unlike MAI-Base-1. The company said the data included image and text pairs, charts, documents, spatial reasoning examples, medical material and computer-use tasks. The model’s knowledge cutoff is March 2026.
Foundry access comes later
Microsoft plans to license the model weights to trusted partners through Foundry. The company said MAI-Base-2 is the foundation for its next generation of models, but it has not announced pricing or a final list of products that will use it.
For now, Microsoft is continuing the post-training process and says it will share more results later.
References
Microsoft AI. (2026, October 15). Frontier pre-training at Microsoft AI. https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/