[X]score: 0.64
Microsoft MAI-Thinking-1: 35B MoE Hits 97% AIME, 53% SWE-Bench Pro
June 2, 2026
Microsoft's MAI-Thinking-1 is a 35B active-parameter MoE model with a 256K context window, scoring 97% on AIME 2025 and 53% on SWE-Bench Pro, placing it alongside Claude Opus 4.6 on coding tasks. Trained without distillation and optimized for Microsoft's MAIA 200 chip, it reportedly runs 30% more efficiently than GB200 hardware. Human raters on Surge preferred it over Claude Sonnet 4.6 in blind evaluations.
HOW THIS AFFECTS YOU
●
builderYou now have a competitive 35B MoE reasoning model with 256K context to evaluate against Anthropic's Sonnet/Opus tier for coding and reasoning workloads.
●
researcherWorth watching because the no-distillation training approach at this scale and the MAIA 200 chip co-design offer a new data point on compute-model co-optimization.
●
founderThis changes the competitive dynamics for reasoning-heavy AI products — Microsoft now has a credible in-house frontier model that could shift Azure pricing and availability.
●
investorMicrosoft fielding a frontier-class model trained on proprietary silicon signals a serious vertical integration play that pressures Anthropic and OpenAI's enterprise positioning.