Microsoft’s MAI-Thinking-1 matches leading reasoning models, trained from scratch

Microsoft announced seven new in-house AI models at Build 2026, headlined by MAI-Thinking-1, the company’s first advanced reasoning model. According to Microsoft, MAI-Thinking-1 is a medium-sized model that “matches leading models” on key software engineering benchmarks. Notably, it was trained entirely from scratch on clean data, without any distillation from third-party models—a deliberate strategy to demonstrate independent AI capability. This launch comes after Microsoft and OpenAI renegotiated their partnership to loosen ties, signaling a shift toward self-reliance in model development.

Alongside MAI-Thinking-1, Microsoft introduced specialized models for various tasks: MAI-Image 2.5 and its flash variant for text-to-image and image editing; MAI-Transcribe-1.5, which is five times faster than competing transcription models; MAI-Voice-2 and a future flash version adding 15 new languages and expanded voice options; and MAI-Code-1-Flash, an inference-efficient coding model already integrated into GitHub Copilot and Visual Studio Code. These models give developers a complete stack of Microsoft-built AI tools, reducing dependence on external providers and offering performance advantages in speed and customization.