Microsoft introduces state-of-the-art MAI Base 2 model with 2T parameters

Microsoft Corp. has developed a new major language model that it claims is significantly more efficient than leading open source alternatives.

The company’s artificial intelligence research group detailed the algorithm on Thursday.

MAI-Base-2 is not a full-fledged LLM, but a basic model, a kind of no-frills prototype. Developers turn base models into LLMs by subjecting them to multiple additional rounds of training. This additional training typically focuses on adding reasoning features and safety guardrails.

The most powerful open source base model currently on the market is DeepSeek-V4-Pro-Base. According to Microsoft, MAI-Base-2 outperformed it on a number of internal benchmarks that included science, programming and math questions. In addition, certain measures make MAI-Base-2 up to 12 times more hardware efficient.

MAI-Base-2 features a mix of Expert or MoE architecture, meaning that only a subset of its artificial neurons are used to process prompts. This is more efficient than activating the entire neural network. MAI-Base-2 has 2 trillion parameters and activates 80 billion of them when it receives a prompt.

The artificial neurons of a language model are organized in clusters called layers. Microsoft’s previous generation base model mixed MoE layers with modules based on a less efficient design called FFN. MAI-Base-2 is made entirely of MoE layers, which contributes to its performance.

Microsoft trained the model for 38 days using 16,000 GB200 chips from Nvidia Corp. The training dataset used in the project included 50 trillion tokens. About a fifth of the tokens were used to train a computer vision module integrated into MAI-Base-2. The module makes the model better than its predecessor at tasks such as understanding scientific diagrams.

Developers train LLMs by observing the way they complete tasks and giving them feedback on how they can improve. This algorithmically generated feedback takes the form of a mathematical object called a gradient. The gradient cycles through the layers of the model being trained and recalibrates each one to improve the output quality.

Gradients change every time they pass through an LLM layer. Under certain circumstances, these changes cause errors that can disrupt the training process. Microsoft has used a technique called gradient clipping to prevent this issue from affecting MAI-Base-2. Muon, a popular open-source AI tool, was also used to speed up the training workflow.

Microsoft will license MAI-Base-2 weights to “trusted partners” via its Azure cloud platform. In addition, the model will be used to power a new series of LLMs. The company plans to release more information about these models in the near future.

Photo: Microsoft