Microsoft's AI Offensive: MAI Models and "Frontier Tuning"

At the Build developer conference, Microsoft announced a major offensive in its own AI models. After the company primarily relied on OpenAI technology for years, the MAI family of AI models is now being created.

At the forefront is Microsoft's frontier model MAI-Thinking-1, which, according to its statement, is its new "Reasoning flagship". The model is intended to compete with rival models in software development and mathematical tasks. According to the company, it was developed entirely with its own training data and not built based on external models.

Microsoft is grouping a whole family of other specialized AI systems around the flagship:

• MAI-Image 2.5 and MAI-Image 2.5 Flash for image generation and editing
• MAI-Code-1-Flash for programming tasks – the model has already been integrated into GitHub Copilot and Visual Studio Code
• MAI-Transcribe 1.5 for speech transcription, which Microsoft claims is up to five times faster than comparable solutions
• MAI-Voice 2 and soon a Flash version for speech synthesis with new voices and 15 languages

For the first time for Microsoft, developers will also be able to adjust the weights of the AIs themselves. These are a kind of focus points that influence an AI's output.

New: AI Learns from Users

Something Microsoft calls "Frontier Tuning" is also new. Here, the AI is supposed to watch the user work and learn from them; for example, how they fill out an Excel spreadsheet. The company praises how great the new "Frontier Tuning" is for companies; the AI would thus perform tasks exactly as they should be done in that company.

However, this should also be very practical for Microsoft. We asked Microsoft to what extent Microsoft is outsourcing the training of its AI agents to end-users and what this means for data protection. The response is still pending.

More news from Build can be found in the Build 2026 overview article.

(rie)