Core Insights - The company "月之暗面" announced the launch of its first Agent product, Kimi-Researcher, which is currently undergoing a small-scale gray testing phase [1] - Kimi-Researcher utilizes end-to-end agentic reinforcement learning technology and has performed competitively in HLE tests, matching the performance of Gemini-Pro's Deep Research Agent [1][4] Product Features - Kimi-Researcher autonomously plans task execution processes and delivers complete results without complex prompts or preset workflows [4] - The model is designed to learn how to think in dynamic environments, making judgments on conflicting information, tool-switching at task nodes, and determining which intermediate information to retain or discard [4] - The primary motivation for Kimi-Researcher is the successful resolution of tasks, emphasizing its focus on task completion [4] Data and Research Integrity - As a deep research model, Kimi-Researcher incorporates a vast array of data sources, allowing for direct traceability of citations, which enhances research rigor and aims to eliminate inaccuracies [4] - The foundational pre-trained model and the model post-reinforcement learning will be gradually open-sourced to promote exploration in the field of agent reinforcement learning [4]
Kimi首个Agent开启小范围灰度测试 性能超OpenAI、Gemini
Huan Qiu Wang·2025-06-21 08:15