MAI-Thinking-1

MAI-Thinking-1 is Microsoft AI’s first large language model and first dedicated reasoning model — a 35-billion-active, roughly 1-trillion-total parameter sparse mixture-of-experts that activates only the slices of the network needed for each request. The model was trained from the ground up on clean data with no distillation from third-party frontier models, end-to-end on Microsoft’s own infrastructure across architecture, pre-training, post-training, reasoning, evaluation, and safety. The 200K-token context window lets it analyze long documents, run multi-step reasoning, and process extended agent traces without chunking, and it supports function calling and developer instructions out of the box.

On hard benchmarks it punches well above its active size — matching Claude Opus 4.6 on SWE-Bench Pro at substantially lower cost, reaching 95.9% on AIME 2025, and earning gold at the 2025 International Math Olympiad. In blind side-by-side evaluations, users rated it at parity with Claude Sonnet 4.6. By owning the model-building loop end-to-end and skipping third-party distillation, Microsoft positions MAI-Thinking-1 as the foundation for enterprise reasoning workloads — complex multi-step instructions, long-context analysis, and code generation — at a lower cost.