Moonshot AI’s open-weights Kimi K2 Thinking model from China surpasses GPT-5 and Claude Sonnet 4.5 across major benchmarks, signalling a global shift where open innovation challenges closed frontier-lab dominance.
China’s Moonshot AI has unveiled Kimi K2 Thinking, an open-weights reasoning model that has outperformed leading closed systems from OpenAI and Anthropic, marking a major inflection point in the evolution of open AI. Released under a modified MIT licence, the model scored 44.9% on Humanity’s Last Exam, surpassing GPT-5 and Claude Sonnet 4.5, and achieved 60.2% on BrowseComp, a key benchmark for agentic search and browsing. It further demonstrated its coding strength with 61.1% on SWE-Multilingual, 71.3% on SWE-bench Verified, and an impressive 83.1% on LiveCodeBench V6.
K2 Thinking excels in agentic tasks requiring long autonomous operations, capable of performing 200–300 sequential tool calls without human input. Built with a 256K context window and designed for test-time scaling, it expands both thinking tokens and tool-calling abilities. The model is currently accessible on kimi.com in chat mode, with full agentic mode coming soon and API access already available.
This performance milestone challenges the long-standing dominance of closed frontier labs, suggesting that open models may now rival proprietary systems in both capability and complexity. By allowing researchers and enterprises to inspect, customise, and deploy AI systems independently, such models could accelerate innovation and reduce dependence on commercial APIs.
Founded in 2023 by Yang Zhilin in Beijing, Moonshot AI’s breakthrough underscores China’s rising influence in global AI. As Western experts including Jensen Huang, Eric Schmidt, and Marc Andreessen, warn of China’s rapid progress, Kimi K2 Thinking’s success hints that the balance of AI power may be shifting sooner than expected.
