Microsoft published a detailed tech report on how they trained their new frontier LLM with reasoning.
TG AI News·June 5, 2026 at 6:25 PM·
Trusted Source
Microsoft published a detailed tech report on how they trained their new frontier LLM with reasoning.
It seems it is not yet on the arena, but the reading is still interesting. It has been a long time since such detailed reports with technical details have been released by Big Tech.
**MAI-Thinking-1** - this is a 35B active / 1T total parameter MoE.
256k token window (fits a 600-page document)
They trained it on a cluster with 8000 GB200.
They focused more on data and training recipes rather than innovations in architecture.
It will not be open-sourced, but they will provide an API for fine-tuning.
[Blog post](https://microsoft.ai/news/introducing-mai-thinking-1/)
[Tech report](https://microsoft.ai/wp-content/uploads/2026/06/main_20260602_2.pdf)