Microsoft AI launches MAI-Thinking-1 reasoning model

Microsoft AI announced MAI-Thinking-1, a sparse mixture-of-experts model with about 1 trillion total parameters and 35 billion active. It is available in public preview through Microsoft Foundry. The company calls it a medium-sized model. It says the model is among the strongest in its weight class and is "toe-to-toe" with Claude Opus 4.6 on SWE-Bench Pro.

Other reported results include 97.0% on AIME 2025 and 94.5% on AIME 2026. In a blind side-by-side human evaluation run with the data-labeling firm Surge across 1,276 single-turn and multi-turn tasks, raters preferred MAI-Thinking-1 over Claude Sonnet 4.6. All of these figures come from Microsoft. The announcement does not give a margin for the preference result. Competitor scores in its benchmark tables are taken from those companies' official model cards.

Microsoft says the model "was trained without distillation from third party models" and uses "clean, traceable, and enterprise-grade" data. It describes the release as part of a "Hill-Climbing Machine," a pipeline meant to keep improving models over time with better data, rewards, training environments and compute. The company also points to in-house infrastructure that runs from its own accelerators to its reinforcement learning framework. For agentic coding, it says it built deterministic, executable training environments graded by real test suites.

The model supports a 256k-token context window, function calling and layered developer instructions, and it works with the Chat Completions API. The page refers to seven new MAI models but does not describe the others in the excerpt provided. It also does not cover pricing.