Recently, Moon's Dark Side announced that its Kimi smart assistant has officially launched the first Agent product — Kimi-Researcher (Deep Research) in a small-scale gray test. This new-generation Agent model trained with end-to-end autonomous reinforcement learning (end-to-end agentic RL) technology is designed to provide users with efficient and in-depth deep research services.

When facing problems, Kimi-Researcher demonstrates strong autonomous planning and execution capabilities. It not only actively clarifies questions and deeply thinks but also autonomously plans keywords for search and screens high-quality information. During task processing, Kimi-Researcher usually performs 23 steps of reasoning, plans 74 keywords, and finds 206 URLs, finally retaining only the top 3.2% of content with the highest information quality. This process ensures the depth and traceability of research results.

WeChat_Screenshot_20250621094225.png

In addition to its powerful information search and screening capabilities, Kimi-Researcher can also autonomously call on tools such as browsers and code to process raw data and automatically generate analysis conclusions. Its deliverables include a detailed and traceable in-depth research report, as well as an interactive and sharable dynamic visualization report. These reports are over ten thousand words long, average about 26 high-quality references, and support online generation of links for sharing, greatly facilitating user presentation and collaboration needs.

To verify Kimi-Researcher's real capabilities, Moon's Dark Side arranged a challenging "exam" for it — Humanity’s Last Exam (HLE). This specially designed high-difficulty benchmark covers hundreds of professional fields, from mathematics, physics, medicine to politics and history, comprehensively testing the model's problem-solving abilities in complex knowledge tasks. Kimi-Researcher achieved excellent scores of 26.9% Pass@1 accuracy and 40.17% Pass@4 accuracy under completely zero-structure and no-process design settings, surpassing several well-known AI models and reaching one of the highest levels known so far.

In real-world applications, Kimi-Researcher also demonstrated outstanding performance. Whether it's algorithm engineers looking for high-value benchmarks, operations staff researching company development within industries, or legal professionals quickly understanding data privacy regulations across countries, Kimi-Researcher can generate structured and comprehensive reports in a short time, providing strong support for users.

Moon's Dark Side stated that Kimi-Researcher is an Agent model trained through end-to-end reinforcement learning, featuring zero-structure and self-adaptation. It has no complex prompts or preset processes but relies entirely on the model's own trial and error and learning to handle complex tasks. This design enables Kimi-Researcher to demonstrate strong adaptability and generalization capabilities when dealing with conflicting information, tool switching, and environmental changes.

Currently, Kimi-Researcher is in the small-scale gray test phase. Users can apply for beta access by visiting kimi.com and start using it by enabling the "Deep Research" button below the Kimi chat box after obtaining permission.