Microsoft unveils seven MAI models, from reasoning to voice and code

Microsoft has announced seven in-house MAI models spanning reasoning, image generation, transcription, voice and coding. The practical story is that some are already available in Microsoft products, while others are arriving through previews or later releases.

Microsoft AI Watch analysis

What happened

At its Build 2026 conference, Microsoft introduced MAI-Thinking-1, a 35-billion-active-parameter reasoning model with a 128,000-token context window. The company says it was built from scratch using commercially licensed data and designed for multi-step instructions, long-context reasoning and code generation. Microsoft also says independent evaluators preferred it to Claude Sonnet 4.6, and that it matched Claude Opus 4.6 on SWE-Bench Pro. Those comparisons are company claims, not independently established results here.

The other announced models cover image generation (MAI-Image-2.5 and a Flash version), transcription (MAI-Transcribe-1.5), voice (MAI-Voice-2 and a Flash version) and coding (MAI-Code-1). MAI-Thinking-1 is in private preview in Microsoft Foundry. MAI-Image-2.5 is already live in PowerPoint and OneDrive, while MAI-Code-1 is available in Copilot and VS Code. Microsoft says the transcription model is coming in 43 languages; its voice models are already available in 15 additional languages.

Why it matters

This is a broad move towards Microsoft-built models inside the tools people already use, not just a research announcement with a model name and a distant promise. The immediate payoff is uneven but tangible: coding and image models are already reaching products, while the flagship reasoning model remains in private preview.

For developers and businesses, Foundry is the route to explore the models as they become available. For everyone else, the key question is whether the models improve everyday work enough to matter, rather than simply expanding Microsoft’s catalogue.

Our read

Seven models give Microsoft a credible-looking spread across common AI tasks. The useful test is not the size of the launch list, but the quality and availability of each model in the product where people need it. A good code assistant is worth more than an impressive family portrait of model names.

What to watch

• Whether MAI-Thinking-1 moves beyond private preview, and on what terms.
• How the image and coding models perform in PowerPoint, OneDrive, Copilot and VS Code.
• Whether Microsoft publishes details that let others assess its benchmark comparisons.

Discussion spark: Would you rather use a strong model built into the Microsoft tools you already rely on, or choose the best model separately for each task?

not affiliated with or endorsed by Microsoft