🤖Microsoft Drops MAI-Thinking-1, Its First Reasoning Model
TL;DR
Microsoft launched MAI-Thinking-1, a 35B-active-parameter reasoning model trained from scratch with no distillation from OpenAI's GPT series. The 1T-parameter MoE hits 97% on AIME 2025 and matches Claude Opus 4.6 on SWE-Bench Pro, signaling Microsoft's bid for model independence.
Microsoft launched MAI-Thinking-1, a 35B-active-parameter reasoning model trained from scratch with no distillation from OpenAI's GPT series. The 1T-parameter MoE hits 97% on AIME 2025 and matches Claude Opus 4.6 on SWE-Bench Pro, signaling Microsoft's bid for model independence.
Key Points
35B active params, ~1T total in a sparse MoE; 256K context window
97.0% AIME 2025 and 94.5% AIME 2026 — matches Claude Opus 4.6 on SWE-Bench Pro
Trained from scratch on commercially licensed enterprise data, no GPT distillation
Available in private preview via Microsoft Foundry, Chat Completions API compatible
Companion MAI-Code-1-Flash turns text descriptions into application source code
Why It Matters
Microsoft now has a frontier reasoning model it owns end-to-end, which weakens Azure's dependence on OpenAI and gives enterprise customers a cheaper second option without leaving the stack.
Quick Facts
Frequently Asked Questions
Why does this matter?
Microsoft now has a frontier reasoning model it owns end-to-end, which weakens Azure's dependence on OpenAI and gives enterprise customers a cheaper second option without leaving the stack.
What happened?
Microsoft launched MAI-Thinking-1, a 35B-active-parameter reasoning model trained from scratch with no distillation from OpenAI's GPT series. The 1T-parameter MoE hits 97% on AIME 2025 and matches Claude Opus 4.6 on SWE-Bench Pro, signaling Microsoft's bid for model independence.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,461 builders reading daily.