# Microsoft opens MAI-Thinking-1 reasoning model in public preview > MAI-Thinking-1 adds an internally trained reasoning model to Microsoft’s lineup, with an enterprise preview and a sparse mixture-of-experts design. Canonical URL: https://www.devobs.io/news/news-mai-thinking-1-preview/ By: Elias Brooks Published: 2026-09-06T11:58:54.629Z Updated: 2026-09-06T11:58:54.629Z Event date: 2026-08-12 Section: AI Microsoft announced MAI-Thinking-1 on August 12 and made the reasoning model available in public preview through Foundry. The [announcement](https://microsoft.ai/news/introducing-mai-thinking-1/) describes a sparse mixture-of-experts architecture and an internally developed training pipeline. ## A new reasoning-model option Microsoft reported approximately one trillion total parameters with 35 billion active parameters, and said the model was trained without distillation from other companies’ models. It presented software-engineering and mathematics evaluations as evidence of its capabilities. Those are vendor-reported characteristics and results. They do not establish that the model will be the strongest choice for every enterprise task, or that a training-data statement settles an individual customer’s governance requirements. ## Evaluate the decision the model must make Our suggested pilot would use tasks with independently checkable outcomes. For a planning assistant, define the constraints that must hold in the final plan. For a calculation workflow, check the result using the application’s own deterministic rules. Keep an expert review path for answers that cannot be verified automatically. Evaluate difficult negative cases too: missing information, contradictory instructions and requests that should lead to a clarification. A model that produces a polished answer despite insufficient inputs may create more review work than one that explicitly asks for the missing fact. ## Preserve a comparable baseline Run the candidate and existing model against the same saved inputs, tools and time limits. Record total interaction cost and completion time alongside correctness, rather than comparing an isolated model score with an end-to-end production workflow. The announcement provides a new candidate for that evaluation. A useful adoption decision should identify which specific class of work improves and how the team will detect regressions after the preview changes. ## Source references -