Microsoft unveils MAI-Base-2, a 2-trillion-parameter base model it says outruns top open-source rivals

MAI-Base-2 is a base model rather than a finished large language model. Developers convert base models into deployable LLMs through additional rounds of training, typically aimed at adding reasoning capabilities and safety guardrails. Microsoft positioned the release as a more efficient starting point for that work.

The model uses a mixture-of-experts architecture, activating 80 billion of its 2 trillion parameters when it receives a prompt rather than running the entire network. Microsoft's previous-generation base model combined mixture-of-experts layers with modules built on a less efficient design known as FFN; MAI-Base-2 consists entirely of mixture-of-experts layers, which the company lists among the contributors to its performance.

Microsoft said MAI-Base-2 beat DeepSeek-V4-Pro-Base, which it described as the most capable open-source base model on the market, on a series of internal benchmarks covering science, programming and mathematics questions. The company also said the model is up to 12 times more hardware-efficient by certain measures. Those comparisons rest on Microsoft's own testing, and the benchmarks were not independently verified in the account.

Training ran for 38 days on 16,000 of Nvidia's GB200 chips using a dataset of 50 trillion tokens. About one-fifth of those tokens went to a computer vision module built into the model, which Microsoft says improves its handling of tasks such as interpreting scientific charts compared with its predecessor.

To keep training stable, Microsoft applied a technique called gradient clipping, which limits the changes that accumulate as gradients pass through the model's layers and can otherwise introduce faults that disrupt a training run. The company also used Muon, a widely adopted open-source AI tool, to speed up the workflow.

Microsoft plans to license the model's weights to "trusted partners" through its Azure cloud platform and will use MAI-Base-2 to power a new series of LLMs. The company said it will share more information about those models in the near future.