MAI-Thinking-1

What it is

MAI-Thinking-1 is the flagship language model of Microsoft AI and its first dedicated reasoning model. A reasoning model is designed to work through a problem step by step before giving an answer, rather than producing a reply in one pass. Microsoft announced it on 2 June 2026 at Microsoft Build, its annual developer conference, as part of a group of seven new MAI models.^[1][2]

It is the successor to MAI-1-preview, the text model Microsoft released earlier.^[1]

Origins

The model came from the same effort that produced Microsoft's earlier in-house models. Microsoft said it was trained from scratch on clean, properly licensed data, without distillation from other labs' models. Distillation here means training a smaller model to copy the outputs of a larger one, and Microsoft said MAI-Thinking-1 was not built that way.^[2]

The model's development also relied on a large compute base. Microsoft said its next-generation GB200 cluster was operational and cited a 1.4x efficiency gain from co-designing models with its in-house Maia 200 chip.^[2]

Main features

Microsoft describes MAI-Thinking-1 as a "medium-sized" model that is top-tier for its class. The company did not publish parameter counts, context length or pricing in the announcement.^[1][2]

On performance, Microsoft said that in blind human side-by-side evaluations the model was preferred to Anthropic's Claude Sonnet 4.6.^[1] It also said the model matches advanced models on major software-engineering benchmarks and shows strong math reasoning.^[2] These are Microsoft's own results and are not independently verified here.

Microsoft also introduced what it called "Frontier Tuning". This means adapting MAI models to specific workflows through reinforcement learning in real customer environments.^[2] Reinforcement learning is a training method in which a model improves by receiving feedback on the results of its actions. Microsoft said MAI models tuned for Excel matched GPT-5.4 with up to 10 times the efficiency.^[2]

A notable change for developers is access. Microsoft said developers could now fine-tune the model weights themselves for the first time, which means adjusting the learned parameters of the model for their own needs.^[2]

Place in the development

MAI-Thinking-1 is available in Microsoft Foundry and in Microsoft products. It is also offered on third-party platforms OpenRouter, Fireworks and Baseten.^[2] Those platforms let developers reach the model without going through Microsoft directly.

In the seven-model launch, the reasoning model led the group, and the other models covered coding, image, transcription and voice.^[1] Within Microsoft AI's own line, it marks the point where the company moved from general language models to a dedicated reasoning system with a public fine-tuning route.^[1][2]