Microsoft introduced MAI-Thinking-1 in public preview on August 12, making its reasoning model available through Microsoft Foundry. The launch announcement emphasizes software engineering, mathematical tasks and enterprise applications that need to combine model responses with tools and internal instructions.
Microsoft describes a sparse mixture-of-experts architecture with about one trillion total parameters and 35 billion active parameters. The active number reflects the part of the model involved in computation, rather than the full size of the model that must be hosted.
The interface supports long context and tools
MAI-Thinking-1 supports a 256K-token context window, function calling and developer instructions. Microsoft says it is compatible with the Chat Completions API, giving application developers a familiar request format.
Compatibility can reduce the amount of integration work needed to try another model. It does not establish that an application’s existing prompts will produce identical behavior. Tool selection, response formatting and instruction handling remain useful areas for comparison.
Microsoft places the model within Foundry’s evaluation, monitoring and deployment environment. The August release is a hosted public preview; the announcement does not offer downloadable weights for local deployment.
That distinction matters for teams considering where their code or documents will be processed. Our guide to files supplied to AI assistants explains why deployment and data-handling arrangements deserve attention alongside a model’s capabilities.
Microsoft’s benchmark results are a starting point
The company reports software-engineering and mathematics results, as well as a blind human comparison with Claude Sonnet 4.6. It says the human evaluation covered 1,276 tasks using professional raters through partner Surge.
Those results come from Microsoft’s evaluation program. A preference result across that task set does not establish which model will perform best on a particular organization’s documents, codebase or workflow.
Microsoft also says it trained the model without distilling from third-party models and used traceable training data. These are claims about its development process, rather than an independent audit reported here.
For a team trying the preview, a useful comparison is a set of its own representative tasks, with consistent inputs and clearly defined success criteria. The context-window size and a published benchmark score answer different questions from whether the model completes those tasks reliably.
