Microsoft Releases MAI-Thinking-1

Microsoft AI has released MAI-Thinking-1, a new reasoning model. This model achieves state-of-the-art (SOTA) performance in its class across mathematics, knowledge, and coding, while providing clean and traceable data suitable for enterprise tasks.

MAI-Thinking-1 is a Sparse Mixture of Experts (MoE) model with 35B-active and approximately 1T-total parameters. It offers efficient inference costs compared to larger models, while demonstrating performance on par with Claude Opus 4.6 on SWE-Bench Pro, making it suitable for the daily workflows of developers and enterprises.

To ensure continuous advancement in model development, Microsoft introduced a co-design pipeline called the Hill-Climbing Machine. This system adheres to the following three philosophies:

• Learned, not Inherited: The model is designed to learn tasks on its own, rather than relying on knowledge distillation from third-party models.
• Clean Data: The model is trained from scratch on enterprise-grade data with clear quality and provenance, allowing for complete understanding and control of its behavior.
• Self-sufficiency: The entire stack is optimized through in-house infrastructure, ranging from custom accelerators to reinforcement learning frameworks.

The model also demonstrates very strong mathematical reasoning capabilities. It recorded 97.0% on AIME 2025 and 94.5% on AIME 2026, proving its excellent mathematical and scientific reasoning performance.