SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Kimi-Researcher - A Deep Research Agent Model Launched by Kimi

Author: AI小集 · Original source: AI工具集

What Is Kimi-Researcher

Kimi-Researcher is a new-generation Agent model launched by Kimi under Moonshot AI, trained on end-to-end agentic RL technology and designed specifically for deep research tasks. It can autonomously plan task execution workflows, and through steps such as clarifying questions, in-depth reasoning, proactive search, and tool invocation, ultimately delivers high-quality research results.

The core capabilities of Kimi-Researcher include: proactively asking follow-up questions to build a clear problem space, averaging 23 steps of reasoning for in-depth thinking, filtering high-quality information through 74 keywords and 206 URLs, and invoking tools to process raw data and generate analytical conclusions. It can output in-depth research reports of more than 10,000 words, cite about 26 high-quality sources, and generate interactive dynamic visualization reports, making it easy for users to quickly grasp the core conclusions.

Kimi-Researcher

Main Features of Kimi-Researcher

  • Clarifying Questions: Through proactive follow-up questions, it helps users build a clearer problem space.
  • In-Depth Reasoning: Each task averages 23 steps of reasoning, autonomously sorting out and resolving requirements.
  • Proactive Search: Plans an average of 74 keywords and filters out the top 3.2% of content with the highest information quality.
  • Tool Invocation: Autonomously invokes tools such as browsers and code to process raw data and generate analytical conclusions.
  • Generating In-Depth Research Reports: Outputs reports of more than 10,000 words citing about 26 high-quality sources, with all citations traceable.
  • Dynamic Visualization Reports: Provides structured formatting and mind maps for quickly grasping core conclusions.
  • Asynchronous Execution: Uses an asynchronous approach to ensure output quality and information coverage.

Technical Principles of Kimi-Researcher

  • End-to-End Agentic RL: Kimi-Researcher adopts an end-to-end reinforcement learning approach. During training, the model autonomously learns through trial and error, treating the entire task as a whole for learning. The model can handle complex reasoning, tool switching, and environmental changes without relying on preset workflows or human-designed prompts.
  • Zero-Structure Design: Kimi-Researcher is a zero-structure Agent, with no complex prompts or preset workflows. The model forms its reasoning patterns on its own during training, and all strategies, paths, and judgments are naturally formed through repeated trial and error.
  • Outcome-Driven Reinforcement Learning Algorithm: The model's only driving force is whether the task is truly solved. Only when the task is completed and the correct result is obtained does the model receive a reward. This ensures that the model can autonomously optimize its behavior when facing complex tasks.
  • Lightweight Long-Term Memory Mechanism: Kimi-Researcher has no fixed memory module; it autonomously decides which information is worth remembering and how to use that information during reasoning. This enables the model to efficiently handle long-sequence tasks.
  • Training Infrastructure for Agents: Kimi-Researcher's training infrastructure supports asynchronous execution and flexible interfaces, and optimizes the learning efficiency of long-sequence tasks through mechanisms such as "step-by-step rollback."
  • Multimodal Capabilities and Long Chain-of-Thought Reasoning: The technical framework of Kimi-Researcher also involves enhancing multimodal capabilities, improving the model's performance in multimodal tasks through training on combined text and visual data. Through long chain-of-thought reasoning training, the model can handle complex logical reasoning tasks.

Kimi-Researcher's project address

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

How to use Kimi-Researcher

  • Access entry: Visit Kimi the official website, or search for "Kimi Smart Assistant" in the WeChat mini program.
  • Apply for beta testing: Or click to apply for beta testing permission, and provide the questions you need Kimi-Researcher to help you research.
  • Use features
    • Deep research: Kimi-Researcher will autonomously plan the task execution process, including clarifying questions, in-depth reasoning, proactive searching, and calling tools, and ultimately generate a deep research report. (20 credits per month, and supports 1 concurrent task at the same time)
    • Dynamic visualization report: Generate a visually structured report with structured layout, making it easy to quickly grasp the core conclusions.
    • Web search: Kimi-Researcher can search the internet for the latest information, integrate and summarize related content.
  • Enter questions or instructions: Enter your questions or specific needs in the dialog box, and Kimi-Researcher will conduct in-depth research according to your instructions.
  • Upload files: Supports uploading files in multiple formats (such as PDF, Word, Excel, PPT, TXT, etc.), with a maximum of 50 files, each no more than 100M.
  • Specify tasks: Clearly tell Kimi-Researcher the operations you need, such as extracting key content, summarizing, translating, etc.
  • Usage tips
    • "Continue" feature: When processing long content, click the "Continue" button to ensure the model maintains coherent thinking.
    • Common phrases feature: Set common phrases or shortcut instructions to quickly trigger specific tasks.
    • Role-playing: Let Kimi-Researcher play a specific role (such as interviewer, expert, etc.) to help complete specific tasks.
  • Verification and validation: For the analysis or conclusions provided by Kimi-Researcher, it is recommended to judge and verify them in combination with your own professional knowledge to ensure the accuracy of the results.

Kimi-Researcher's benchmark tests

  • "Humanity's Last Exam" (Humanity’s Last Exam, HLE)：
    • Pass@1 accuracy：26.9%
    • Pass@4 accuracy：40.17%
    • This performance exceeds Claude 4 Opus (10.7%), Gemini 2.5 Pro (21.6%), is slightly higher than OpenAI Deep Research (26.6%), and is tied with Gemini-Pro's Deep Research Agent (26.9%).
  • Sequoia China xbench benchmark test：In the DeepSearch task, Kimi-Researcher's average pass rate reached 69%, leading other models on this leaderboard.

Application scenarios of Kimi-Researcher

  • Real-time research support: Users can ask about the latest research progress, and Kimi will search and provide related papers, data, and analysis reports.
  • Market trend analysis: Analyze market trends, consumer behavior, and competitor strategies, and provide detailed market analysis reports.
  • Lesson Plan Writing: Teachers can use Kimi-Researcher to write lesson plans and generate complete teaching structures.
  • Legal and Government Affairs Scenarios: Automatically identify risk clauses and generate revision suggestions. The evidence chain is automatically organized and matched with legal provisions to generate a case summary report with legal basis.