Microsoft launches MAI‑Thinking‑1, its first home‑grown AI that reads 600 pages and tops rivals

Microsoft showed its new AI model MAI-Thinking-1 at the Build 2026 conference. The model is the first advanced reasoning system built entirely inside Microsoft. It leads a new family of MAI models that the company announced at the same event.

MAI-Thinking-1 uses a Sparse Mixture‑of‑Experts design. It runs with 35 billion active parts and about one trillion total parts. It can look at up to 256 thousand words at once, enough for very long documents.

The model scored 52.8 percent on the software‑engineering test SWE‑Bench Pro, matching top rivals. It got 97.0 percent on the math test AIME 2025, 94.5 percent on AIME 2026, and 87.7 percent on LiveCodeBench v6. In blind tests with more than 1,200 tasks, professional raters chose it over Claude Sonnet 4.6 for quality, help, and following instructions.

Microsoft trained the model from scratch on clean, licensed human data. No parts were copied from other companies’ models. The training used a repeatable system called the Hill‑Climbing Machine, which links data pipelines, reinforcement‑learning loops, evaluations, and a fast YOLO training framework.

The company markets MAI‑Thinking‑1 as a medium‑size, cost‑effective model for business use. A private preview runs on Microsoft Foundry and a public preview will appear on the MAI Playground soon, both with strong security and compliance. The model is one of seven new MAI tools that cover coding, images, transcription, voice and more, and will be added to products like GitHub Copilot and Azure through platforms such as Microsoft Foundry, OpenRouter, Fireworks and Baseten.