Kimi K2 Thinking raises the bar for open-source

Alibaba-backed Kimi K2 Thinking, an open-source thinking model that meets or surpasses models like GPT-5 and Claude 4.5 Sonnet across a number of benchmarks at far cheaper cost, was recently released by the Chinese startup Moonshot AI. The specifics: Kimi achieved a new top score of 44.9% on Humanity's Last Exam and outperformed GPT-5 and Sonnet 4.5 on a number of agentic benchmarks. Although it currently lags behind the best models, the model also demonstrated significant coding gains over its predecessor from just four months ago. K2 Thinking excels in creative writing and can independently link together 200–300 tool calls to complete tasks. According to reports, the model's training cost was less than $5 million, and its price is far less than that of the existing frontier models. Jensen Huang of Nvidia stated a few days ago that China is "nanoseconds" behind in AI. This undoubtedly reflects that sentiment, whether or not he was alluding to Moonshot. With price that makes it a very real rival to the best closed solutions, K2 Thinking is the closest Chinese and open-source labs have come to the frontier.