Microsoft unveils MAI-Base-2, a 2-trillion-parameter base model

Microsoft says its MAI-Base-2 base model outperforms DeepSeek-V4-Pro-Base on internal science, programming and math benchmarks and is up to 12 times more hardware-efficient by certain measures.

The model has 2 trillion parameters, activates 80 billion per prompt, and was trained for 38 days on 16,000 Nvidia GB200 chips. Microsoft will license its weights to trusted partners through Azure and use it to power a new series of LLMs.