MAI-Thinking-1 (microsoft.ai)

Microsoft has unveiled MAI-Thinking-1, its new medium-sized reasoning model designed to excel in software engineering tasks and advanced mathematical reasoning. With 35 billion active parameters, this model not only matches leading counterparts, like Claude Opus 4.6, on critical benchmarks but also outperforms them in human evaluations. The model was built from scratch using clean, enterprise-grade data rather than distillation from third-party sources, emphasizing a philosophy of true learning over imitation. This approach aims to enhance its adaptability and utility in real-world applications. Alongside the model, Microsoft introduced the Hill-Climbing Machine, a systematic pipeline for iterative model improvement that ensures each phase of development is optimized for better performance. This initiative addresses the need for cleaner data, self-sufficient training infrastructure, and continuous learning from user interactions. Notably, MAI-Thinking-1 is tailored for enterprise use, featuring a long context window suitable for complex documents and an alignment with core business needs. Its dual focus on capabilities and safety balances advanced functionality while adhering to ethical standards. As MAI-Thinking-1 advances towards Humanist Superintelligence, it reaffirms Microsoft's commitment to creating AI that serves humanity, ensuring tech remains under human control.