Microsoft AI finishes pre-training MAI-Base-2, a 2.8T-parameter mixture-of-experts model

MAI-Base-2 has 80 billion active and 2.8 trillion total parameters. Microsoft trained it on 16,000 GB200 GPUs over 38 days in August and September, with average model FLOPs utilization of 21% and goodput of 89%. The company says it beat every open-weight model it tested on held-out evaluations and claims 4 to 12 times better pre-training FLOPs efficiency than DeepSeek-V4-Pro-Base, savings it put into training a larger model. Kimi K3 and Qwen3.8-2.4T-A95B were also compared, though only post-trained checkpoints exist for them. The evaluations are Microsoft's own. No weights or release date are public, and the model is still in post-training. Microsoft says pre-training used organic data without distillation from other models, which it says will let it license weights to trusted partners through Foundry. The last release was MAI-Thinking-1 in June, built on MAI-Base-1. For founders, a frontier-scale supplier with clean data provenance could matter to enterprises wary of distilled models, and licensed weights through Foundry would be a new way to buy frontier capability outside the usual labs.