Microsoft AI introduces seven new in-house models grouped under the MAI family.

Microsoft AI launches seven MAI models, including MAI-Thinking-1 and MAI-Code-1-Flash, trained on Maia 200 chips to target humanist superintelligence.

Designed as a multimodal ecosystem, these models cover reasoning, code, image, transcription, and voice. The flagship reasoning model, MAI-Thinking-1, according to the company, ranks among the best in its size class and would be preferred over Sonnet 4.6 in its blind human evaluations. Alongside it are MAI-Code-1-Flash, an agentic code model with five billion parameters integrated into GitHub Copilot and VS Code, MAI-Image-2.5 for image generation and editing, MAI Transcribe-1.5, which supports 43 languages, and MAI-Voice-2, capable of reproducing a voice from a short sample in about fifteen languages. Microsoft emphasizes its method: all are trained from scratch, without distillation from third-party models, on data it presents as clean and licensed, and co-designed with its in-house silicon Maia 200. Distributed via Foundry, developers will for the first time be able to adjust their weights themselves. On this occasion, the company details its Microsoft Frontier Tuning approach, based on reinforcement learning in environments specific to each organization, so that a model learns directly from internal workflows while remaining the property of the client. In this same vein, a partnership with the Mayo Clinic aims to co-create a health model that will remain owned by the hospital. All of this aligns with the stated goal of Humanist Superintelligence, i.e., systems designed to remain tools under human control.