Microsoft Unveils MAI-Base-2 MoE Model

Microsoft recently completed pre-training MAI-Base-2, an MoE foundation model featuring 80 billion active and 2.8 trillion total parameters. Trained on 16k GB200 GPUs over 38 days in August and September 2026, the model claims 4x to 12x more pre-training FLOPs-efficiency than DeepSeek-V4-Pro-Base. It outperforms every open-weight model tested on held-out evaluations, with the company planning to license weights through Foundry.