Kimi K2 Thinking introduces a new open-source thinking model that excels in reasoning and problem-solving. This model performs well on various benchmarks, achieving a score of 44.9% on Humanity's Last Exam (HLE) using tools, which includes search and web-browsing capabilities. It also scored 60.2% on BrowseComp and 71.3% on SWE-Bench Verified, showcasing its strong generalization abilities. K2 Thinking can execute between 200 to 300 sequential tool calls without human help, enabling it to tackle complex problems through coherent reasoning across multiple steps. The model’s strengths lie in its agentic reasoning and coding capabilities. It demonstrated outstanding performance in competitive programming tasks and problem-solving scenarios. For instance, it solved a PhD-level mathematics problem that involved a complex sampling procedure in hyperbolic space, using 23 interleaved reasoning and tool calls. The problem required evaluating a function related to a probability density function (pdf) of a random variable. K2 Thinking's approach involved breaking down the problem, exploring various mathematical transformations, and applying concepts from hyperbolic geometry. K2 Thinking's functionality extends to real-time information collection and agentic search, enhancing its usefulness for coding and analytical tasks. The model's architecture supports deep reasoning and the use of diverse tools, making it suitable for both academic inquiries and practical applications. With its full agentic mode expected to launch soon, K2 Thinking represents a significant advancement in open-source AI models designed for complex reasoning tasks.