Microsoft has a new foundation to build on. Its AI research group unveiled MAI-Base-2 on Thursday, a base model the company says outperforms the strongest open alternatives on science, coding and math questions.
The scale is serious. MAI-Base-2 carries 2T parameters and wakes 80B of them per prompt through a mixture-of-experts design, the architecture that keeps giant models affordable to run.
What changed is the recipe. Microsoft’s previous base model mixed expert layers with older feed-forward blocks. The new one is experts all the way through, and the company credits that uniformity for part of the gain. Training took 38 days across 16,000 Nvidia GB200 chips and 50 trillion tokens.
Roughly a fifth of those tokens went to a vision module folded into the model, which sharpens its reading of scientific charts and diagrams. Microsoft measures up to 12 times better hardware efficiency than its own earlier system on some tests.
Treat the numbers carefully. The comparisons come from Microsoft’s internal benchmarks rather than a neutral leaderboard, and a base model is a raw prototype, not a finished assistant. Developers still have to add the instruction tuning and safety work that turns it into something users touch.
Still, the framing is pointed. Microsoft named DeepSeek-V4-Pro-Base, the best open base model available, as the bar to clear. Beating it on efficiency would matter to anyone paying for GPU hours.
