Microsoft unveils MAI-Base-2 with 2 trillion parameters

Microsoft's AI research group has detailed MAI-Base-2, a 2-trillion-parameter base model that activates 80 billion parameters per prompt using a mixture-of-experts architecture. The company says it outperforms DeepSeek-V4-Pro-Base on internal benchmarks covering science, programming, and math, and is up to 12 times more hardware-efficient. Trained over 38 days on 16,000 Nvidia GB200 chips using 50 trillion tokens, it also includes a computer vision module. Microsoft plans to license the model's weights to trusted partners via Azure and use it to power a new LLM series.