Amidst the ever-growing landscape of AI assistants, a new entrant has emerged from the shadowy realm of Moon's Dark Side: Kimi-Researcher. This innovative tool, currently in the midst of a limited beta test, promises to revolutionize the way we approach deep research through its end-to-end autonomous reinforcement learning technology.
Kimi-Researcher stands out with its impressive autonomous planning and execution capabilities. It takes a proactive approach to clarifying questions and diving into deep contemplation, autonomously planning keyword searches, and filtering out only the highest quality information. On average, Kimi-Researcher engages in 23 steps of reasoning, identifies 74 keywords, scours 206 websites, and retains only the top 3.2% of information quality, ensuring the depth and traceability of its research findings.
Beyond its formidable search and filtering prowess, Kimi-Researcher also autonomously utilizes browsers and code to process raw data and generate analytical conclusions. It delivers comprehensive, traceable deep research reports alongside interactive, shareable dynamic visualizations. These reports, which exceed ten thousand words in length, average about 26 high-quality sources and support the generation of shareable online links, greatly facilitating user presentation and collaboration needs.
To verify Kimi-Researcher's capabilities, Moon's Dark Side subjected it to a rigorous "Humanity’s Last Exam" (HLE), a benchmark designed for AI that spans hundreds of professional fields, from mathematics and physics to politics and history. In a completely unstructured, process-free setting, Kimi-Researcher achieved a Pass@1 accuracy of 26.9% and a Pass@4 accuracy of 40.17%, surpassing several well-known AI models and reaching one of the highest levels known to date.
In real-world applications, Kimi-Researcher has proven its mettle, whether it's helping algorithm specialists find valuable benchmarks, aiding operational colleagues in studying industry company development, or assisting legal professionals in quickly understanding data privacy regulations across different countries. It generates clear, comprehensive reports in a short span, providing robust support to users.
Moon's Dark Side highlights that Kimi-Researcher is an Agent model trained through end-to-end reinforcement learning, characterized by its zero-structure and adaptability. It relies entirely on the model's trial and error and learning to tackle complex tasks without complex prompt words or preset processes. This design allows Kimi-Researcher to demonstrate strong adaptability and generalization capabilities when facing information conflicts, tool switching, and environmental changes.
Currently, Kimi-Researcher is in the phase of limited gray testing. Interested users can apply for beta access at kimi.com and, upon obtaining permissions, initiate their deep research journey by clicking the "Deep Research" button below the Kimi chat interface.