Microsoft: from small Phi models to in-house MAI models

A tech veteran that also builds its own models

Founded in 1975 and based in Redmond, near Seattle, Microsoft is best known in AI for its partnership with OpenAI and its Copilot assistant. But the company also develops its own models, with two very different approaches: research into small, efficient models (Phi) and, since the creation of the Microsoft AI division led by Mustafa Suleyman, an in-house family meant to compete at the top (MAI).

Phi: the power of small models

Phi is Microsoft’s family of small language models. The idea is to get strong results from a compact model trained on carefully curated data, rather than from a giant one. Versions have followed one another (Phi-3, Phi-3.5, then Phi-4 with its mini, multimodal and reasoning-focused variants).

Their key strength is the license: Phi models are released as open weights under the permissive MIT license. You can get them from Hugging Face, through Ollama or in Azure AI Foundry. In practice, they suit on-device use: an app on a laptop, a device without a constant connection, or an internal tool that must not send data outside the company.

MAI: the Microsoft AI models

The MAI family is newer. It includes a reasoning model (MAI-Thinking), a lightweight coding model built for GitHub Copilot and VS Code (MAI-Code), plus models for image generation (MAI-Image), speech synthesis (MAI-Voice) and transcription (MAI-Transcribe). Microsoft frames the effort as “humanist”: AI designed to help people and organizations.

Unlike Phi, Microsoft does not present MAI as downloadable models. They are distributed through an API on Microsoft Foundry and partner platforms such as OpenRouter, Fireworks and Baseten, with customization options for businesses.

What makes Microsoft different

Microsoft plays on several fronts at once: it hosts other labs’ models on its cloud, develops its own closed models and releases small open ones. This mixed approach gives you plenty of choice, but it can be confusing: depending on the Microsoft product you use, the model answering you is not always the same.

To find your way, compare models in our comparison tool, browse the open-source models list or read the profile of OpenAI, Microsoft’s long-standing partner.