Kimi-Researcher - Deep Research Agent Model launched by Kimi

What is Kimi-Researcher?

Kimi-Researcher is part of Dark Side of the Moon. Kimi Introducing a new generation of Agent models trained based on end-to-end agentic RL technology, designed for deep research tasks. It can autonomously plan the task execution process and deliver high-quality research results through clarifying the question, in-depth reasoning, active searching, and invoking tools. It actively asks questions to build a clear problem space, averages 23 steps of reasoning for in-depth thinking, filters high-quality information through 74 keywords and 206 URLs, and invokes tools to process raw data and generate analytical conclusions.

Key Features of Kimi-Researcher

• Points of clarification: Help users construct a clearer problem space by initiating rhetorical questions.
• inference: An average of 23 steps of reasoning per task to autonomously sort out and address needs.
• Proactive search: 74 keywords were planned on average to filter out the top 3.2% content with the highest quality of information.
• invocation tool: Autonomously invoke tools such as browsers and codes to process raw data and generate analysis conclusions.
• Generate in-depth research reports: Output of reports of more than 10,000 words, citing some 26 high-quality sources, with all citations traceable.
• Dynamic visualization reports: Structured layout and mind maps are provided to facilitate a quick grasp of the core conclusions.
• asynchronous execution: Asynchronous approach is used to ensure output quality and information coverage.

Kimi-Researcher Program Address

• Technical Papers:: https://moonshotai.github.io/Kimi-Researcher/

How to use Kimi-Researcher

• access portal: Access Kimi Official website, or search "Kimi Intelligent Assistant" in WeChat mini program.
• Apply for internal testing: or click Apply for beta access to provide a question that Kimi-Researcher can help you with.
• Usage Functions
  • in-depth studyKimi-Researcher will autonomously plan the task execution process, including clarifying questions, in-depth reasoning, active searching and calling tools, and finally generating in-depth research reports. (20 quota per month, supports 1 task concurrently)
  • Dynamic visualization reports: Generate visual reports with structured layouts for quick and easy access to core findings.
  • Internet search: Kimi-Researcher searches the Internet for the latest information, consolidates and summarizes relevant content.
• Enter a question or instruction: Enter your question or specific needs in the dialog box and Kimi-Researcher will conduct in-depth research based on your instructions.
• Uploading files: Supports uploading files in multiple formats (e.g. PDF, Word, Excel, PPT, TXT, etc.), up to 50 files, each file not exceeding 100M.
• mandate: Tell Kimi-Researcher exactly what you need to do, such as extracting key content, summarizing, translating, and so on.
• Tips for use
  • "Continue" function: Clicking the "Continue" button when working with long content ensures that the model stays coherent.
  • common language function: Set common phrases or shortcut commands to quickly trigger specific tasks.
  • role-playing (as a game of chess): Have Kimi-Researcher play a specific role (e.g., interviewer, expert, etc.) to help accomplish a specific task.
• Verification and CalibrationKimi-Researcher recommends that any analysis or conclusions provided by Kimi-Researcher be judged and verified by Kimi-Researcher in conjunction with its own professional knowledge to ensure the accuracy of the results.

Kimi-Researcher's Core Strengths

• Strong autonomous planning and execution capabilities: Kimi-Researcher can autonomously plan the task execution process, proactively clarify questions, build a clearer problem space, and filter high-quality information through multi-step reasoning (23 steps on average) and keyword planning (74 keywords on average).
• Efficient information search and filtering: In each task, Kimi-Researcher visits an average of 206 URLs, filtering out the top 3.2% content with the highest quality of information.
• In-depth research reports and dynamic visualizations: Kimi-Researcher offers two deliverables:
  • In-depth research report: The average length is more than 10,000 words, with citations from about 26 high-quality sources, all embedded in the text and clickable to jump to and highlight the original article.
  • Dynamic visualization reports: Supports structured layout and mind mapping to make important information clear at a glance, and online link generation and sharing.
• Excellent benchmark performanceIn the Humanity's Last Exam (HLE), a difficult benchmark test designed for AI, Kimi-Researcher achieved Pass@1 accuracy of 26.9%, Pass@4 accuracy of 40.171T, and TP3T. TP3T, surpassing several well-known AI models. In Sequoia China's xbench benchmark test, Kimi-Researcher achieved an average pass rate of 69% in the DeepSearch task, ahead of other models.
• Zero Structure and Adaptive Capabilities: Kimi-Researcher is an Agent model trained through end-to-end reinforcement learning without complex cue words or predefined processes. It can autonomously adapt to complex tasks, showing strong generalization ability and adaptability.
• Asynchronous execution and high quality output: To ensure output quality and information coverage, Kimi-Researcher utilizes an asynchronous execution approach to progressively reason, retrieve and author content. This ensures the high quality and detail of the final research and visualization reports delivered.

Who is Kimi-Researcher for?

• Academic researchers: It can quickly read a large amount of literature, extract key information, and generate a detailed literature review report. Supports simultaneous analysis of Chinese and English literature, helping researchers quickly understand the international cutting-edge research dynamics. Provide services such as concept explanation, viewpoint comparison, grammar checking, sentence rewriting, etc. to help improve the quality of papers.
• Business Analyst: Can analyze market trends, consumer behavior and competitors' strategies and generate detailed market analysis reports. Provide company financial reports, industry analysis and market forecasts to help users make investment decisions. Quickly sort out the strengths and weaknesses of competing products and generate comparative analysis reports to assist business decisions.
• educator: Helps teachers write lesson plans to generate a complete teaching structure, including modules for introduction, author introduction, content comprehension, and classroom exercises.
• legal practitioner: Automatically identify risky clauses and generate amendment suggestions to help legal staff quickly review contracts. Automatically combing the chain of evidence, matching legal provisions, and generating summary reports to assist in case analysis.
• Technology Developer: Quickly read API documents , generate code , Debug , support for multiple programming languages . Provide error analysis, solutions and related code examples to help developers quickly solve problems.