Microsoft AI has completed the pre-training of MAI-Base-2, a state-of-the-art Mixture-of-Experts (MoE) foundation model boasting 2.8 trillion total parameters and 80 billion active parameters. Trained on 16,000 NVIDIA GB200 GPUs over 38 days, the model demonstrates exceptional data efficiency, proving 4 to 12 times more FLOPs-efficient than leading open base models like DeepSeek-V4-Pro-Base across held-out evaluations in code, math, STEM, and QA. The architecture introduces several key advancements, including native vision-language integration, untied embeddings, optimized Muon optimizers, and a transition to MoE layers across the entire network. Leveraging robust scaling ladders and strict data provenance using organic training tokens without distillation, Microsoft plans to license MAI-Base-2 weights to trusted enterprise partners through its Foundry platform.

Full report : The newly pre-trained Mixture-of-Experts model outperforms major open-weight checkpoints while pioneering advanced scaling and hardware co-design.