Kimi-Researcher – Kimi’s deep research Agent model

What is Kimi-Researcher?

Kimi-Researcher is a next-generation Agent model developed by Moonshot AI under its Kimi brand. It is trained using end-to-end agentic reinforcement learning (RL) technology and is designed specifically for in-depth research tasks. The agent autonomously plans task execution workflows, clarifies questions, performs deep reasoning, actively searches for information, and uses external tools to deliver high-quality research results.

Core Capabilities of Kimi-Researcher

• Clarifies Questions: Engages in counter-questioning to help users define a clearer problem space.

• Deep Reasoning: Performs an average of 23 reasoning steps per task to deeply analyze and resolve the query.

• Active Search: Strategically plans 74 keywords on average and filters down to the top 3.2% of high-quality information.

• Tool Use: Automatically calls on browsers, code execution tools, and others to process raw data and generate insights.

• Generates Long-Form Research Reports: Delivers reports exceeding 10,000 words, citing around 26 high-quality and traceable sources.

• Interactive Visual Reports: Outputs structured layouts and mind maps to help users grasp core insights quickly.

• Asynchronous Execution: Uses asynchronous workflows to ensure comprehensive information coverage and high output quality.

Technical Principles Behind Kimi-Researcher

• End-to-End Agentic Reinforcement Learning: Trained as a unified task-solving agent through trial and error, Kimi-Researcher learns without relying on predefined workflows or prompt engineering.

• Zero-Structure Design: Operates without preset prompts or templates. The reasoning patterns are learned autonomously through reinforcement over time.

• Outcome-Driven RL Algorithm: The model is only rewarded when a task is genuinely completed with correct outcomes, ensuring it optimizes for real results.

• Lightweight Long-Term Memory: Instead of using a fixed memory module, the model decides which information to retain and how to retrieve it during reasoning.

• Agent-Oriented Training Infrastructure: Supports asynchronous execution, flexible APIs, and mechanisms like “stepwise rollback” to boost long-sequence learning.

• Multimodal & Long-Chain Reasoning: Trained to integrate text and visual data for complex tasks and execute long reasoning chains effectively.

Project Page

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

How to Use Kimi-Researcher

• Access Point: Visit the official Kimi website or search “Kimi智能助手” on WeChat Mini Programs.

• Apply for Beta: Submit a research question to apply for access to Kimi-Researcher.

Available Functions

• Deep Research: Kimi-Researcher autonomously plans and executes workflows including question clarification, deep reasoning, search, and tool usage. (Up to 20 tasks/month, 1 concurrent task supported)

• Dynamic Visual Reports: Provides structured and interactive visual reports for quicker understanding.

• Web Search: Can fetch and summarize the latest online information.

• Input Questions or Commands: Simply enter your query or instructions to initiate a research task.

• File Upload: Supports up to 50 files in formats like PDF, Word, Excel, PPT, TXT (max 100MB each).

• Task Assignment: Directly instruct Kimi-Researcher to extract key points, summarize, translate, etc.

Usage Tips

• “Continue” Feature: Keeps reasoning coherent when processing long content.

• Quick Commands: Set up reusable phrases for faster task initiation.

• Role Play: Let Kimi-Researcher act as an interviewer, expert, etc., for specific scenarios.

• Validation: Cross-check Kimi’s outputs using your expertise to ensure accuracy.

Kimi-Researcher Benchmarks

• Humanity’s Last Exam (HLE):

  • Pass@1 Accuracy: 26.9%

  • Pass@4 Accuracy: 40.17%
    This performance outperforms Claude 4 Opus (10.7%) and Gemini 2.5 Pro (21.6%), slightly surpasses OpenAI Deep Research (26.6%), and is on par with Gemini-Pro’s Deep Research Agent (26.9%).

• Sequoia China xBench Benchmark:

  • Achieved a 69% average pass rate on DeepSearch tasks, ranking at the top among all tested models.

Application Scenarios for Kimi-Researcher

• Real-Time Research Support: Ask about the latest developments, and Kimi will return relevant papers, data, and reports.

• Market Trend Analysis: Analyze market trends, consumer behavior, and competitor strategies to deliver comprehensive reports.

• Lesson Planning: Educators can use Kimi-Researcher to create complete lesson plans and instructional frameworks.

• Legal & Government Use: Automatically identifies risky clauses and generates amendment suggestions; organizes evidence chains and matches legal provisions to produce case summaries with legal references.