Chinese Kimi Researcher surpassed giants in Humanity’s Last Exam

We have another technological breakthrough from China! The Kimi.ai project presents research agent Kimi Researcher. Which surpasses flagship solutions of technology giants.

Kimi Researcher demonstrates truly impressive results! The agent confidently leads in all key benchmarks. Including the prestigious Humanity’s Last Exam. A competition considered the gold standard for evaluating AI research capabilities.

The technical implementation impresses with thoughtfulness. When performing one research task, the system executes an average of 23 reasoning steps and simultaneously analyzes over 200 unique URLs. This is an unprecedented level of analysis depth that ensures maximum accuracy and completeness of results.

Special attention deserves the fine-tuning of tool usage with application of improved multi-agent dialogue methods. We’re talking about the MCP protocol. This allows the agent to effectively combine various research strategies and adapt to complex information flows.

The basic foundation of the agent became their own language model k1.5. Which the Kimi team subjected to end-to-end additional training using reinforcement learning methods.