Microsoft AI has finished pretraining MAI-Base-2, a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active, trained on 50 trillion tokens across 16,000 Nvidia GB200 GPUs over 38 days. The claim attached to it is a 4x to 12x advantage in pretraining FLOPs efficiency over DeepSeek-V4-Pro-Base.
Two things to hold about that number. It is a pretraining-efficiency claim, which is about how much compute was needed to reach a given capability — not a statement about how the finished model performs against anything. And the comparison is against a base model, so it is comparing like with like at the pretraining stage rather than against a shipped product.
MAI-Base-2 is still in post-training. When it ships it goes to trusted partners through Foundry, not as open weights. So unlike Beam or Nemotron, this is not a model anyone outside that programme will be running.
The strategic read is the same one as Maia 300 earlier this week: Microsoft is building the full stack — its own silicon, its own frontier model, its own serving platform — and reducing how much of Azure's AI business depends on other people's components.
Sources:
Microsoft AI ·
Techmeme