MAI-Thinking-1 — benchmark results

Microsoft AI's first in-house reasoning model, a sparse MoE (1T total, 35B active). Provider: Microsoft. Released 2026-06-02. Access: API.

Unified ELO 1642 ± 38, rank #403 of 1841 rated models, from 8 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
LLM Stats (AIME 2026)94.5Score (%)75
ZeroEval GPQA Diamond84.2GPQA Diamond Score74.1
LLM Stats (LongBench v2)61Score (%)66.7
LLM Stats (Multi-Challenge)53Score (%)46.4
MedXpertQA43Score (self-reported)44
BenchLM50.4Overall Score42.6
LLM Stats (HMMT Feb 26)84.9Score (%)21.4
LLM Stats (HealthBench Professional)35Score (%)0

Interactive version: theaggregate.ai/model?slug=mai-thinking-1 · How It Works · Data refreshed daily, snapshot 2026-07-25.