Skip to content
Microsoft AIPreview

MAI-Thinking-1

Microsoft’s in-house reasoning flagship — frontier ambitions on MAIA silicon.

Read the full MAI-Thinking-1 analysis

Context

256K

Max output

—

Input /1M

—

Output /1M

—

Best for

  • Watching Microsoft’s full-stack frontier play (chip + model + tuning)
  • Math/reasoning workloads (vendor-claimed 97% AIME 2025)
  • Frontier Tuning into custom company-specific agents

Watch out

Vendor-claimed at launch — human-rater preference over Sonnet 4.6 and SWE-Bench Pro figures are unreproduced. Await LMArena/ARC/Artificial Analysis before trusting in production.

For creators. One to track, not yet to standardize on. Re-evaluate once independent benchmarks land.

Benchmarks

aime 202597
swe bench pro53

Capabilities

  • Text reasoning foundation model
  • 35B-active mixture-of-experts
  • 256K context window
  • Co-designed for Microsoft MAIA 200 silicon
  • Microsoft Frontier Tuning (custom company-specific agents)

More from Microsoft AI

Microsoft Phi-4 (open-weight family)MAI-Image-2.5MAI-Code-1-Flash

Sources