Microsoft Debuts MAI-Thinking-1, Its First In-House Reasoning Model
Microsoft AI (@MicrosoftAI) unveiled MAI-Thinking-1, its first full-fledged in-house reasoning model, on June 2 during Build 2026. Built from scratch as a 35B active / ~1T total MoE model, it claims performance competitive with similar-sized models on STEM reasoning and coding tasks. The model scored 52.8% on the coding benchmark SWE-Bench Pro, matching Claude Opus 4.6, and reached 97% on the math reasoning benchmark AIME 2025 (Microsoft AI post, model page).
Details were released simultaneously on the company's official site and on X, and the model became available in Microsoft Foundry's private preview. MAI-Thinking-1 is positioned as one of seven new MAI models spanning reasoning, coding, image, voice, and transcription (announcement blog).
The announcement marks a strategic turning point for Microsoft. The company had largely relied on OpenAI's models but began accelerating its own model development around 2025. MAI-Thinking-1 emphasizes being built "from scratch," with "zero distillation from third-party models" and "clean and appropriately licensed data" that excludes AI-generated content, foregrounding data sovereignty and enterprise trust. The Microsoft AI team led by Mustafa Suleyman frames it as the first step of a "hill-climbing machine" that automates continuous improvement, linking it to Frontier Tuning—RL adaptation on enterprise workflows—and a healthcare collaboration with Mayo Clinic (announcement blog, The Verge).
On specifications, it uses a sparse MoE architecture with a 35B active / ~1T total parameter configuration designed to keep inference costs down. Its context length is reported at 256K tokens (some reports cite 128K), and in Surge's blind side-by-side evaluation it surpassed Sonnet 4.6 on overall quality (Microsoft News). It launched immediately in Microsoft Foundry's private preview, with a public preview and availability on MAI Playground, Fireworks AI, and Baseten planned soon. Baseten will support fine-tuning with weight ownership and eyes-off data (Microsoft AI, Baseten). The model also features built-in safety guardrails and copyright protection.
On X, the official posts spread actively and drew strong engagement. Developers praised the combination of enterprise fine-tuning and RL environments as a strength, while others noted it is currently available only via Foundry's private preview, with no local or open-weight support. Technical threads praised its "persistence of thinking" that holds up over thousands of steps and debated how its approach differs from DeepSeek and GLM-5 (reaction 1, reaction 2). Still, as a preview-stage release, reports of large-scale real-world use cases remain limited (Neowin).