Microsoft Announces MAI-Thinking-1 and MAI-Code-1-Flash Models
- •Microsoft released MAI-Thinking-1, a 35B parameter reasoning model, and MAI-Code-1-Flash, a 5B parameter coding model.
- •MAI-Thinking-1 reportedly outperforms Sonnet 4.6 in internal blind human side-by-side performance evaluations.
- •Both models were trained using exclusively clean and commercially licensed data, excluding third-party model distillation.
Microsoft released two new text-focused large language models (LLMs) on June 2, 2026, aimed at distinct reasoning and coding applications. The first model, MAI-Thinking-1, features 35B parameters and is designed for complex reasoning tasks. Microsoft reports that the model is preferred over Sonnet 4.6 in internal blind human side-by-side evaluations. Access is currently limited to select early partners.
The second release, MAI-Code-1-Flash, is a 5B parameter model optimized specifically for developers. Microsoft developed this model to integrate into GitHub Copilot and Visual Studio Code, focusing on delivering high performance at a lower cost compared to larger alternatives. The company stated that the model is beginning its rollout to individual GitHub Copilot users.
Both models were trained from the ground up using clean and commercially licensed data, explicitly avoiding distillation from third-party models. This development highlights a shift toward models built on carefully vetted datasets rather than large-scale web scrapes.