Microsoft used its Build 2026 conference to launch seven in-house MAI models across reasoning, coding, image generation, transcription, and voice. The lead reasoning model, MAI-Thinking-1, has 35 billion parameters. Microsoft said its blind human evaluations put it on par with Claude Sonnet 4.6, while coding benchmarks matched Claude Opus 4.6. The company also reported 97% on AIME and 53% on SWE Bench Pro.
The lineup includes MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5 and its Flash version, MAI-Transcribe-1.5, MAI-Voice-2, and the upcoming MAI-Voice-2-Flash. Microsoft said the models will be available through Azure Foundry and its own product stack, and will also be listed on third-party platforms including OpenRouter, Fireworks, and Baseten. The benchmark claims, though, were presented by Microsoft itself, with no independent third-party validation yet available.
Enterprise customization sits at the center of the rollout
Beyond benchmark claims, Microsoft placed heavy emphasis on Frontier Tuning. The pitch is straightforward: enterprises can use records from their own workflows to train a customized version of an MAI model, while the training data and the resulting model weights remain owned by the customer rather than flowing back to Microsoft.
The system relies on a Reinforcement Learning Environment, or RLE, which Microsoft described as a simulated training ground built around real business tasks. In company-provided examples, an MAI model tuned for Excel workflows reached parity with GPT 5.4 and delivered up to a 10x efficiency gain. A version tuned to McKinsey enterprise standards posted the highest blind-test win rate with costs about 10x lower. Those figures also came from Microsoft’s own disclosures.
Mayo Clinic project highlights a customer-owned AI model
Microsoft also pointed to its work with Mayo Clinic as a concrete example of the strategy. The two sides are building a frontier medical AI model using Mayo’s clinical expertise and de-identified clinical data, while full ownership of the model remains with Mayo. The system is set to be validated internally at Mayo first, then later made available to other healthcare institutions through Azure Foundry.
That structure positions Microsoft as the infrastructure layer rather than the ultimate owner of the AI system. For large organizations, the distinction matters. It affects data control, deployment boundaries, and how much legal and operational authority remains with the customer.
Zero distillation is also a compliance message
Throughout the announcement, Microsoft repeatedly said the MAI models were trained from scratch, without distilling any third-party models and without using unauthorized or unclear-source data. The company framed the training pipeline as clean and fully licensed.
That message carries both a technical and legal angle. It separates Microsoft from the distillation-heavy approach referenced in the source material, which noted that Chinese labs such as DeepSeek have used distillation to reproduce near-top-tier performance at much lower cost. At the same time, the clean-data claim speaks directly to enterprise buyers worried about copyright lawsuits and regulatory exposure. In that context, licensed data and zero distillation may matter as much as benchmark numbers.
From the Build presentation, Microsoft’s bigger bet appears to be less about topping public model rankings and more about selling a framework where customers build and own their own AI systems. The remaining question is when outside evaluations will test the company’s published performance claims.