June 21, 2011 - 1AI has learned from theDark Side of the Moon Kimi Publicis has learned that Kimi's first Agent（Agent）Kimi-Researcher Opens on the 20thlocalGrayscale test.
Kimi-Researcher is a new generation of agent model trained based on end-to-end agentic RL technology, and is also an agent product specially designed for deep research tasks. Later, Dark Side of the Moon will gradually open source Kimi-Researcher's basic pre-trained models and reinforcement learning models.
For each problem, Kimi-Researcher autonomously plans the task execution process and ultimately delivers complete results:
- Clarification: Proactive rhetorical questioning in understanding the problem and constructing a clearer problem space;
- reflect in depth: An average of 23 steps of reasoning per task to autonomously sort out and address needs;
- Proactive searchFor each task, an average of 74 keywords were planned, 206 URLs were found, and the top 3.2% contents with the highest information quality were judged and filtered by the model to eliminate redundant and low-quality information;
- Invoke the tools and deliver the results: Autonomously invoke the browser, code and other tools to process raw data, automatically generate analysis conclusions, and complete the delivery end-to-end.
To ensure the quality of the output and the coverage of the information, Kimi-Researcher uses theasynchronous executionway, spending more time progressively reasoning, retrieving and writing content.
Users will eventually receive 2 deliverables.
An informative and traceable in-depth research report
- The average length of a report is betweenMore than 10,000 words;
- An average of about 26 high-quality, traceable sources are cited;
- All citations are embedded in the body of the text, clickable to jump and highlighting the original text for easy verification and traceability.
An interactive, shareable and dynamic visualization report
- Structured layout and mind maps make trends, exceptions and other important information visible at a glance;
- The overall structure and core conclusions can be quickly grasped without having to read the entire text;
- Support online link generation and sharing for easy display.
It's official, Kimi-Researcher has been recognized as one of the most successful players in the Humanity's Last Exam (HLE), a difficult benchmark designed specifically for AI, and has been recognized as one of the best in the world.Zero structure, no flow designThe scores for the settings are as follows:
- Pass@1 Accuracy: 26.9%
- Pass@4 Accuracy: 40.17%
This performance exceeds Claude 4 Opus (10.71 TP3T), Gemini 2.5 Pro (21.61 TP3T), is slightly higher than OpenAI Deep Research (26.61 TP3T), and is tied with Gemini-Pro's Deep Research Agent (26.91 TP3T), which This is one of the highest levels known to date. In the xbench benchmark released by Sequoia China, an AI capability evaluation system aligned with real task scenarios, Kimi-Researcher achieved an average pass rate of 69% in the DeepSearch task, ahead of other models in the list.