Model
mai-thinking-1
Appears alongside
Featured in threads
Tracks
- Models & capabilities 2
- Compute & infrastructure 1
- Labs & people 1
Microsoft AI details MAI-Base-2, a 2.8-trillion-parameter model trained in-house
Trained on 16k GB200 GPUs over 38 days, it was compared on held-out bits-per-byte with open-weight models only; Microsoft plans to licence the weights through Foundry.
Models & capabilities · Compute & infrastructure
Microsoft launches in-house MAI-Thinking-1 reasoning model at Build 2026
Trained from scratch on licensed data with no distillation from OpenAI or any other lab, the sparse model activates 35bn of roughly 1 trillion parameters per query.
Models & capabilities · Labs & people