Microsoft has completed the pre-training of a 2.8-trillion-parameter AI model

Microsoft has completed pre-training of its new artificial intelligence model, MAI-Base-2. According to the company, the model outperformed every open-weight model tested, while requiring 12 times less compute to achieve results comparable to a competing model.

MAI-Base-2 has a total of 2.8 trillion parameters, though it activates only 80 billion of them for any given task. This approach, known as the Mixture-of-Experts architecture, helps make more efficient use of computational resources.

According to Microsoft, across four different evaluations, MAI-Base-2's training process was 4.2 to 12.3 times more efficient in terms of compute usage compared to DeepSeek-V4-Pro-Base. In addition, compared to the previous model, MAI-Base-1, the efficiency of the training methodology increased by 6.3 to 12.7 times.

To train the model, Microsoft used 16,000 Nvidia GB200 GPUs over 38 days in August and September. The training data comprised approximately 50 trillion tokens. The number of unique tokens increased by about 36% compared to MAI-Base-1, which was released in June.

The company stated that the model was trained on original data rather than through distillation — the method of transferring knowledge from other AI systems. This is an important distinction for Microsoft, as the company is considering licensing the model's weights to trusted partners through Azure AI Foundry in the future.

Microsoft attributes the improved results to advances in data quality, model architecture, and training infrastructure. The company also incorporated visual data into the new model's training process from the outset — image-related material accounted for about 18% of the training data.

The release of MAI-Base-2 marks another step in Microsoft's strategy to build its own advanced AI models and reduce reliance on external providers. The company is developing its own models for integration into its products and for enterprise customers.

At this stage, Microsoft has not announced a commercial availability date or pricing for MAI-Base-2. The model is still undergoing additional training and optimization, while the company is already preparing to develop MAI-Base-3.