Last night, Dark Side of the Moon officially released its new generation open-source thinking model, Kimi K2 Thinking. Trained based on the "model as Agent" concept, this model natively possesses the ability to "think while using tools," achieving state-of-the-art (SOTA) performance in multiple benchmark tests, including Humanity's Last Exam and BrowseComp.
According to the official introduction, Kimi K2 Thinking can perform up to 300 rounds of tool calls and stable multi-round thinking without human intervention, autonomously solving complex problems. In Humanity's Last Exam, covering over 100 professional fields, it achieved a new SOTA score of 44.9%. In the BrowseComp test released by OpenAI, it achieved an accuracy rate of 60.2%, far exceeding the human average of 29.2%, demonstrating extremely strong information mining and reasoning capabilities.
This model uses a dynamic loop of "thinking → searching → programming" to decompose fuzzy problems into executable tasks. For example, in a stock buyback case, it identifies the target company through two rounds of searching and extracts data from the SEC website, ultimately providing an accurate conclusion. Furthermore, its coding capabilities have improved in multilingual engineering benchmarks (SWE-Multilingual, etc.), and it performs better in terminal operation and long text processing.
In terms of general capabilities, Kimi K2 Thinking can transform inspiration into coherent narratives in creative writing, deeply analyze complex instructions in academic research, and demonstrate greater empathy in emotional responses. The model is currently available on Hugging Face and ModelScope platforms, and developers can deploy and use it for free.
