AB
AB
AiBoss

All Categories

project

Kimi-Researcher - Kimi's deep research agent model

Kimi-Researcher, developed by Kimi under the Dark Side of the Moon, is a next-generation agent model trained using end-to-end agentic reinforcement learning (RL) technology, designed specifically for deep research tasks. It can...

What is Kimi-Researcher?

Kimi-Researcher, developed by Kimi under the Dark Side of the Moon, is a next-generation agent model trained using end-to-end agentic reinforcement learning (RL) technology, designed specifically for deep research tasks. It can autonomously plan its task execution process, delivering high-quality research results through steps such as clarifying the problem, in-depth reasoning, proactive searching, and invoking tools.

Kimi-Researcher's core capabilities include: proactively asking questions to construct a clear question space; conducting an average of 23 steps of reasoning for in-depth analysis; filtering high-quality information through 74 keywords and 206 URLs; and utilizing tools to process raw data and generate analytical conclusions. It can output in-depth research reports exceeding 10,000 words, citing approximately 26 high-quality sources, and generating interactive, dynamically visualized reports to help users quickly grasp the core conclusions.

Kimi-Researcher's main functions

  • Clarify the issueBy proactively asking questions, we help users build a clearer question space.
  • Deep reasoningEach task involves an average of 23 steps of reasoning, allowing the system to independently analyze and resolve requirements.
  • Active searchOn average, 74 keywords were planned, and the top 3.2% of content with the highest information quality were selected.
  • Call toolsIt can autonomously call upon tools such as browsers and code to process raw data and generate analytical conclusions.
  • Generate in-depth research reportsOutput a report of over 10,000 words, citing approximately 26 high-quality sources, all of which are traceable.
  • Dynamic Visualization ReportIt provides structured layout and mind maps to facilitate quick understanding of core conclusions.
  • Asynchronous executionAsynchronous operation is used to ensure output quality and information coverage.

The technical principles of Kimi-Researcher

  • End-to-end autonomous reinforcement learningKimi-Researcher employs an end-to-end reinforcement learning approach, where the model learns autonomously through trial and error during training, treating the entire task as a whole. The model can handle complex reasoning, tool switching, and environmental changes without relying on pre-defined processes or human-designed prompts.
  • Zero-structure designKimi-Researcher is a zero-structure agent, without complex prompts or pre-defined processes. The model develops its own reasoning patterns during training; all strategies, paths, and decisions are naturally formed through repeated trial and error.
  • Outcome-driven reinforcement learning algorithmsThe model's sole driving force is whether the task is actually solved. The model is only rewarded when the task is completed and the correct result is obtained. This ensures that the model can autonomously optimize its behavior when faced with complex tasks.
  • Lightweight long-term memory mechanismKimi-Researcher does not have a fixed memory module; it autonomously decides which information is worth remembering and how to retrieve it during inference. This enables the model to efficiently handle long sequence tasks.
  • Agent-oriented training infrastructureKimi-Researcher's training infrastructure supports asynchronous execution and flexible interfaces, optimizing the learning efficiency of long sequence tasks through mechanisms such as "step-by-step rollback".
  • Multimodal capabilities and long-chain reasoningKimi-Researcher's technical framework also involves enhancing multimodal capabilities. By training with combined text and visual data, it improves the model's performance in multimodal tasks. Through long thought chain reasoning training, the model can handle complex logical reasoning tasks.

Kimi-Researcher project address

  • Technical Papershttps://moonshotai.github.io/Kimi-Researcher/

How to use Kimi-Researcher

  • Access pointVisit the official Kimi website or search for "Kimi Smart Assistant" in WeChat mini programs.
  • Apply for internal testingAlternatively, click to apply for beta testing access and provide the questions you want Kimi-Researcher to research for you.
  • Use Function
    • In-depth researchKimi-Researcher autonomously plans the task execution process, including clarifying the problem, in-depth reasoning, proactive searching, and utilizing tools, ultimately generating an in-depth research report. (20 tasks per month, supporting one concurrent task)
    • Dynamic Visualization ReportGenerates structured, visually appealing reports for quick and easy understanding of key conclusions.
    • Internet searchKimi-Researcher can search for the latest information online, integrate and summarize relevant content.
  • Enter a question or instructionEnter your question or specific requirements in the dialog box, and Kimi-Researcher will conduct in-depth research based on your instructions.
  • Upload FileIt supports uploading files in various formats (such as PDF, Word, Excel, PPT, TXT, etc.), with a maximum of 50 files that can be uploaded, and each file must not exceed 100MB.
  • Designated task: Clearly tell Kimi-Researcher what operations you need, such as extracting key content, summarizing, translating, etc.
  • Usage Tips
    • "Continue" functionWhen processing long content, click the "Continue" button to ensure the model maintains a coherent flow of thought.
    • Commonly used phrases functionSet frequently used phrases or shortcuts to quickly trigger specific tasks.
    • role playKimi-Researcher can play a specific role (such as an interviewer or expert) to help complete a specific task.
  • Verification and validationFor analyses or conclusions provided by Kimi-Researcher, it is recommended to use your own professional knowledge to judge and verify them to ensure the accuracy of the results.

Kimi-Researcher benchmarks

  • "Humanity's Last Exam" (HLE):
    • Pass@1 Accuracy26.9%
    • Pass@4 Accuracy40.17%
    • This performance surpasses Claude 4 Opus (10.7%) and Gemini 2.5 Pro (21.6%), slightly exceeds OpenAI Deep Research (26.6%), and is on par with Gemini-Pro's Deep Research Agent (26.9%).
  • Sequoia China xbench benchmark:In DeepSearch tasks, Kimi-Researcher achieved an average pass rate of [percentage missing]. 69%It outperforms other models on the list.

Application scenarios of Kimi-Researcher

  • Real-time research supportUsers can inquire about the latest research progress, and Kimi will search and provide relevant papers, data, and analysis reports.
  • Market Trend AnalysisAnalyze market trends, consumer behavior, and competitor strategies to provide detailed market analysis reports.
  • Lesson plan writingTeachers can use Kimi-Researcher to create lesson plans and generate complete teaching structures.
  • Legal and political scenariosAutomatically identifies risky clauses and generates revision suggestions. The chain of evidence is automatically analyzed and matched with legal provisions to generate a case summary report with legal basis.
Back