Microsoft’s AI research group has built a base model it says outperforms the best open-source equivalents while using a fraction of the compute intensity.
MAI-Base-2 packs two trillion parameters in a mixture-of-experts architecture, activating roughly 80 billion of them per prompt. Base models are the raw material developers later train into useful assistants, and the closest comparison point is DeepSeek-V4-Pro-Base.
Weights locked behind partner licences
On internal benchmarks covering science, programming and maths, Microsoft says MAI-Base-2 came out ahead, and in some measures it was up to 12 times more hardware-efficient. The design change was structural: the previous generation mixed expert layers with conventional feed-forward modules, while the new model uses expert layers throughout.
Training took 38 days on 16,000 Nvidia GB200 chips over 50 trillion tokens, with about a fifth of that budget going to a vision module that helps the model read scientific charts. Microsoft used gradient clipping and the open-source Muon optimiser to keep the run stable.
The distribution plan is the interesting part for startups. Rather than releasing the model openly, Microsoft will license the weights to trusted partners through Azure and use the base to power a new family of language models it plans to detail soon.
That leaves builders weighing an efficient but gated model against freely downloadable rivals from DeepSeek and Alibaba.
