Microsoft AI has introduced a family of seven new models built internally, along with a broader strategy centered on agentic AI, enterprise tuning and specialized deployment. The announcement, published by Microsoft AI chief Mustafa Suleyman, positions the company’s model work as part of a longer-term effort to build systems that can improve continuously as compute and training scale up.
The models span reasoning, coding, image generation and editing, transcription, and speech. Microsoft says the collection is intended to function as a multimodal stack that can support practical work across business and consumer settings. The company also said the models will be distributed through Microsoft’s own platform and made available to developers through third-party services including OpenRouter, Fireworks and Baseten.
At the center of the release is MAI-Thinking-1, which Microsoft describes as its flagship reasoning model. The company says it performed strongly on software engineering and math tasks and was preferred in internal blind comparisons against a competing model. Microsoft also said it trained the system from scratch using clean data rather than distilling it from other labs’ models.
MAI-Code-1-Flash is aimed at coding workflows and is built for use across GitHub Copilot, VS Code and other Microsoft products. Microsoft says the model is designed to be efficient and compact, with about 5 billion active parameters.
Other additions include MAI-Image-2.5, which handles text-to-image generation and image editing, and MAI-Transcribe-1.5, which Microsoft says offers strong transcription performance, multilingual support and fast processing. MAI-Voice-2 focuses on speech generation, with support for 15 languages and safeguards meant to reduce misuse. Microsoft also pointed to lower-cost or more efficient variants of several models, including flash versions that are expected later.
The company said developers will be able to tune model weights themselves for the first time, a move it framed as part of a push toward greater customization. Microsoft also emphasized that the models were trained on what it calls clean, traceable and enterprise-grade data.
Beyond model releases, Microsoft highlighted a method it calls Frontier Tuning. The approach uses reinforcement learning in real-world environments so models can adapt to specific workflows inside an organization. Microsoft likens the environments to training spaces where models learn from the traces of completed work, including steps, decisions and actions.
The company argues this gives businesses more control over how AI behaves and makes it possible to build systems trained on an organization’s own data. Microsoft said an internally tuned Excel model matched GPT 5.4 while using up to 10 times less compute. It also said early enterprise users have seen similar efficiency gains.
This focus on tuning reflects a broader shift in Microsoft’s messaging, with the company framing AI less as a generic assistant and more as a system that can be customized for specific jobs.
Microsoft also announced a collaboration with Mayo Clinic to build a frontier AI model for healthcare. The model will use Mayo Clinic’s clinical expertise, de-identified data and longitudinal insights along with Microsoft’s AI technology. Microsoft says the goal is to support clinical reasoning and healthcare use cases that current general-purpose systems cannot address as well.
The model will first run within Mayo Clinic’s own environment. If validated, Microsoft says it could later be offered to other organizations through Microsoft Foundry. Ownership of the model will remain with Mayo Clinic, which the two organizations say is intended to support trust, safety and responsible data stewardship.
Microsoft framed the announcement as part of a larger buildout of its AI lab, including its own Maia 200 silicon and a focus on internally developed infrastructure. The company says the long-term goal is to build systems that continuously improve while remaining under human control.