全能视觉助手来了!Kimi低调上线K2.5,可同时调度100个智能体,效率最高提升4.5倍
Hua Er Jie Jian Wen·2026-01-27 06:57

Core Insights - The company has made significant upgrades to its flagship model, launching the Kimi k2.5, which enhances its multi-modal capabilities and agent cluster collaboration, aiming to solidify its leading position in the competitive Chinese AI market [1][3]. Model Features - The Kimi k2.5 model utilizes a native multi-modal architecture, capable of processing text, images, and videos simultaneously through a single prompt. Its most notable breakthrough is the "agent cluster" capability, allowing it to autonomously schedule up to 100 sub-agents to work in parallel, reducing execution time for complex tasks by up to 4.5 times [1][3]. - The model has been pre-trained on approximately 15 trillion mixed visual and text tokens, enabling deep visual understanding beyond simple OCR text recognition. Users can upload complex diagrams or financial reports, and K2.5 can analyze and derive logic from them [5]. Agent Cluster Technology - The core highlight of this update is the "agent cluster" paradigm, which employs Parallel Agent Reinforcement Learning (PARL) technology. This allows K2.5 to manage and orchestrate a cluster of 100 sub-agents without predefined roles or workflows, significantly enhancing its ability to handle large-scale searches and complex workflows [7]. - Internal testing indicates that this parallel processing mechanism reduces end-to-end runtime by 80% and supports up to 1,500 coordinated steps in parallel workflows, breaking through traditional single-agent capability limitations [7]. Performance Metrics - Kimi k2.5 has outperformed open-source counterparts in several benchmark tests, particularly in programming and logical reasoning. It aims to compete with top proprietary models, such as Claude Code from Anthropic PBC, through its automated coding tools [9]. - In practical office scenarios, K2.5 has shown a 59.3% performance improvement in handling end-to-end tasks like document processing and financial modeling compared to its predecessor, K2 Thinking [9]. Financing and Market Context - The company, founded by former Tsinghua University professor Yang Zhilin, has recently completed a funding round, raising $500 million from investors including Alibaba and IDG Capital, with a post-money valuation of $4.3 billion. It is currently seeking to raise funds at a valuation of up to $5 billion [3][10]. - The competitive landscape is intensifying, with rivals like Zhipu and MiniMax Group Inc. recently raising over $1 billion through IPOs. The launch of K2.5 ahead of competitors is a strategic move to demonstrate the company's ongoing technological advancement and capital attractiveness [10].