The Microsoft Build 2026 conference unveiled seven new AI models, and the self-developed inference model MAI-Thinking-1 made its debut.

At the Build 2026 developer conference, Microsoft unveiled seven new AI models all at once, covering inference, coding, vision, speech, and multi-modal full-stack capabilities. The flagship inference model MAI-Thinking-1 achieved parity with Anthropic Claude Sonnet 4.6 in the coding benchmark test, and the 5-billion-parameter coding model MAI-Code-1-Flash targeted the GitHub ecosystem. This is Microsoft’s first self-developed advanced inference model, trained from scratch using full and clean data for medium-sized projects, without using distilled data, and achieving industry-leading performance in software engineering benchmark tests. Also released at the same time were MAI-Image 2.5 (including the Flash version), MAI-Transcribe-1.5 (with a speed five times that of competitors), MAI-Voice-2 (supporting 15 languages), and MAI-Code-1 (already integrated into GitHub Copilot and VS Code). This move marks Microsoft’s comprehensive expansion of its self-developed MAI model family, covering inference, image, speech, and programming, and is also an important step for it to reduce its reliance on OpenAI

.