It’s no longer just about having the biggest AI model; it’s about who controls the raw materials needed to create them. At its annual Build conference, Microsoft AI introduced a new, in-house ecosystem of artificial intelligence models. This marks a shift away from industry shortcuts and a commitment to in-house research focused on self-sufficiency in tech. The new MAI family includes seven different models that cover text reasoning, coding, transcription, voice generation, and image processing.

These models are built entirely from scratch, using proprietary licensed data. They come at a key moment for the company. Previously, Microsoft depended almost entirely on systems developed by OpenAI.

At the heart of this announcement is a flagship reasoning model called MAI-Thinking-1. Designed to compete in the medium-sized weight class, this system represents Microsoft’s most ambitious step into advanced reasoning without relying on third-party foundations.

The company claims that this model has matched leading commercial rivals on key software engineering tests and achieved human preference parity with Anthropic’s Claude Sonnet 4.6 in blind evaluations. Alongside the flagship reasoning model, Microsoft is rolling out specialized tools that integrate directly into its software. This includes MAI-Code-1-Flash for GitHub Copilot and Visual Studio Code, MAI-Image-2.5 for text-to-image editing, and MAI-Transcribe-1.5, which works five times faster than competing market models.

Voice generation is managed by MAI-Voice-2, which supports 15 languages and can mimic natural speech from brief audio clips. Importantly, Microsoft believes that the future of enterprise AI will depend on deep customization instead of general-purpose algorithms. With a process called “Frontier Tuning,” clients can now train these foundational models in private digital environments. Early corporate use suggests that these focused systems can outperform larger models.