Microsoft AI launches seven MAI models led by MAI-Thinking-1

Background

Microsoft AI had already released a series of in-house models before this announcement. By June 2026 the company's line included speech, text, image and transcription models, and it was ready to present a larger group of them at once.^[1]

What happened

On 2 June 2026, at Microsoft Build, Mustafa Suleyman announced seven new in-house MAI models. He did so in a post titled "Building a hill-climbing machine".^[1] The models were MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Image-2.5-Flash, MAI-Transcribe-1.5, MAI-Voice-2 and MAI-Voice-2-Flash. The last of these was listed as coming soon.^[1]

Microsoft put MAI-Thinking-1 at the head of the group. It described the flagship reasoning model as preferred to Claude Sonnet 4.6 in its blind human evaluations.^[1]

Key details

MAI-Code-1-Flash is an agentic coding model with 5 billion active parameters. It is built for GitHub Copilot and Visual Studio Code, and Microsoft said it is comparable to Claude Haiku at lower cost.^[1][2] The "active parameters" figure counts the parameters used for each token in a mixture-of-experts model. An agentic coding model works through multi-step programming tasks, not just single answers.

MAI-Transcribe-1.5 supports 43 languages, and MAI-Voice-2 generates speech in 15 languages.^[1] For comparison, the first transcription model, MAI-Transcribe-1, had launched in April 2026 with 25 languages.

The group was offered in Microsoft Foundry, and on the third-party platforms OpenRouter, Fireworks and Baseten.^[2] Offering the models through these platforms lets developers use them without a direct contract with Microsoft.

Microsoft also announced a collaboration with Mayo Clinic to build a healthcare AI model. Mayo Clinic will own the resulting model.^[2] That ownership term is significant, because it means the clinic, not Microsoft, holds the model that comes out of the work.

Impact

The announcement showed Microsoft AI's model work spreading across several tasks at once: reasoning, coding, image, transcription and speech. It also widened the company's reach beyond consumer and developer tools into a specialised area of medicine.^[2]

For readers, the event is a snapshot of where Microsoft's in-house models stood in mid-2026. The claims come from Microsoft's own announcement, and independent evaluation of each model would be needed before drawing firm conclusions about how it compares with rival systems.^[1][2]