MAI-Thinking-1 is a medium-sized reasoning model developed by Microsoft AI, introduced on June 2, 2026, as part of Microsoft's broader initiative towards Humanist Superintelligence. The model is a 35B-active, approximately 1T-total parameters, sparse Mixture of Experts (MoE) model, designed to offer a smaller inference footprint while matching leading models in its weight class. It was trained from the ground up on enterprise-grade, clean, and commercially licensed data, without distillation from third-party models.

On key software engineering benchmarks, MAI-Thinking-1 performs at the level of Claude Opus 4.6 on SWE-Bench Pro. This is particularly relevant for developers and enterprises, as its smaller inference footprint facilitates deployment in daily workflows. The training for agentic coding involved heavily invested, deterministic, executable environments graded by real test suites, allowing the model to practice multi-step coding tasks including reading, editing, running tests, observing failures, and recovering from mistakes. In mathematical reasoning, MAI-Thinking-1 achieves 97.0% on AIME 2025 and 94.5% on AIME 2026, indicating strong capabilities in this domain. This performance is attributed to a training loop that cultivates reasoning gains from proprietary data, rewards, and evaluation processes, enabling generalization.

User preference evaluations, conducted through blind human side-by-side tests with professional raters from Surge across 1,276 tasks, showed MAI-Thinking-1 to be preferred over Claude Sonnet 4.6. These evaluations focused on helpfulness and goal advancement, measuring traits such as task understanding, instruction following, level of detail, clarity, and efficiency. This human preference data is a core focus of post-training, ensuring benchmark improvements translate into better user experiences.

The core methodological innovation underpinning MAI-Thinking-1 and future models is the "Hill-Climbing Machine," a co-designed pipeline engineered to make every component of model development "climbable," ensuring continuous and reliable improvement. This system is designed to absorb better data, stronger rewards, more capable environments, and increased compute. The philosophy is guided by three main pillars:

1. Learned Capabilities: Intelligence is learned directly, not inherited or distilled. MAI-Thinking-1's training without distillation from third-party models forced it to truly learn tasks, which provides greater steerability essential for real-world usage, unlike imitators tied to a teacher's design choices.
2. Clean Data: The model was trained exclusively on clean, traceable, and enterprise-grade data. This emphasis ensures quality, clear provenance, and control over the model's behavior, which is crucial for understanding and improving its outputs.
3. Self-Sufficiency Across the Entire Stack: Efforts are concentrated on in-house training infrastructure, ranging from co-design of models with Microsoft's own accelerators to the proprietary reinforcement learning framework. This end-to-end control allows for full optimization and shaping of systems to best serve specific needs.

MAI-Thinking-1 is built for enterprise readiness, supporting a 256k token context window (equivalent to a 600-page document), function calling, and flexible developer instructions. Its alignment to enterprise needs is achieved through multi-layer instruction following and a default style. Compatibility with the widely used Chat Completions API and integration with Microsoft Foundry for enterprise-grade security and compliance further enhance its commercial viability.

The overarching vision is Humanist Superintelligence, aiming for advanced AI that serves people and organizations without replacing them. This mandates that models remain subordinate technologies under human control, upholding human autonomy and helpfulness. The delicate balance between helpfulness and safety is addressed by treating "unsafe compliance and unnecessary refusal as defects." Safety is trained using the same reinforcement learning infrastructure as capability, where safety rewards are part of the same hill-climbing loop. Reward construction aggregates based on the severity of potential harm, ensuring safety is inherently aligned with capabilities and not an incidental add-on, allowing the model to balance safety bars on sensitive requests with helpfulness on non-sensitive content.

MAI-Thinking-1 is currently available in private preview on Microsoft Foundry, with a public preview on MAI Playground anticipated soon.