On October 15th, Microsoft AI announced the completion of pre-training for its next-generation Hybrid Expert (MoE) foundation model, MAI-Base-2, with a total of 2.8 trillion parameters and 80 billion activation parameters. The model was trained for 38 days from August to September using approximately 16,000 NVIDIA GB200 GPUs. Microsoft stated that on its internally retained test set, MAI-Base-2 outperformed all open-weight models tested. Using DeepSeek-V4-Pro-Base as a benchmark, and estimating the computational cost required to achieve the same pre-training loss level, Microsoft's new training method achieves a 4.2 to 12.3x improvement in floating-point efficiency, and a 6.3 to 12.7x improvement compared to MAI-Base-1 released in June. Microsoft stated that the new model uses its own processed training data, without using distilled data from other models, and for the first time natively integrates visual capabilities during the pre-training phase. Currently, MAI-Base-2 is still in the post-training stage, and Microsoft plans to license the model weights to trusted partners through a foundry in the future; however, a formal release date has not yet been announced.
