Microsoft’s AI research team announced MAI-Base-2, a 2 trillion‑parameter base model. It uses a mixture‑of‑experts (MoE) architecture that activates only 80 billion neurons per prompt, making it up to 12 times more hardware‑efficient on internal tests. The model was trained for 38 days on 16,000 Nvidia GB200 chips with a 50‑trillion‑token dataset, including a vision component. Microsoft will license the model’s weights to trusted partners via Azure and will build downstream LLMs on top of it.
Why it matters
Developers who get access via Azure could build more capable AI services while using less compute power.
