MAI-Thinking-1: Building a Hill-Climbing Machine [pdf] (microsoft.ai)

Microsoft AI has announced MAI-Thinking-1, a groundbreaking model developed from scratch that harnesses a hill-climbing machine approach for continuous improvement in AI capabilities. This 35 billion active parameter model, known for its strong performance on STEM reasoning and coding benchmarks, was pre-trained using 30 trillion tokens from high-quality, enterprise-level data sources, eschewing third-party model distillation or synthetic data. The focus on clear, validated data and scientific rigor aims to provide more robust and steerable AI intelligence. The significance of MAI-Thinking-1 for the AI/ML community lies in its innovative methods for training and reinforcement learning, emphasizing iterative model development as a systemic optimization challenge. The model excels in various evaluations, achieving impressive scores on tasks like SWE-BenchPro (52.8%), AIME2025 (97.0%), and LiveCodeBenchv6 (87.7%). Furthermore, by integrating safety and helpfulness training during the reinforcement learning process, Microsoft aims to balance user interaction with security. This model not only sets new performance standards but also establishes a transparent framework for future AI development, underscoring the importance of human-centric approaches in machine learning advancements.