Microsoft unveils seven in-house MAI models and Foundry agent layer at Build 2026

Microsoft spent two days at Build 2026 arguing it no longer wants to rent the core of its AI business. The centerpiece, presented by Microsoft AI chief Mustafa Suleyman, was seven in-house MAI models spanning reasoning, coding, image generation, voice and transcription—all trained from scratch on licensed data with no distillation from rival labs. The flagship reasoning model, MAI-Thinking-1, uses a sparse mixture-of-experts design with roughly 35 billion active parameters. New silicon tuned for agents and a next-generation quantum chip rounded out the 'ownership' theme across Windows, Azure and GitHub.

On the platform side, Microsoft announced general availability of Work IQ APIs (launching June 16) for building agents over Microsoft 365 data, plus Foundry guided guardrails, enhanced tracing and agent governance, with the knowledge layer reorganized around Foundry IQ unifying Work IQ, Fabric IQ, Azure SQL and Web IQ. The Agent 365 SDK moved to GA and a new GitHub Copilot desktop app extends Copilot into managing tasks and pull requests. Microsoft 365 Copilot SKUs become permanent on July 1 at $23.50 and $32 per user monthly.

Suleyman framed this not as a break from OpenAI but as 'self-sufficiency,' saying the company was 'set free' to pursue superintelligence. Competitively, it positions Microsoft against both its own partner OpenAI and against Google and Anthropic. Early hands-on reviews were mixed: a PCMag tester found the four new models uneven against established frontier systems, suggesting MAI is a credible hedge rather than a frontier-resetting leap. Suleyman separately touted MAI-Transcribe-1.5 as best-in-class on Artificial Analysis benchmarks. Watch whether enterprises actually switch Copilot's backend from OpenAI/Anthropic to MAI.