Moonshot AI has released Kimi K2.5, which it describes as the most powerful open-source model to date. Building on the previous K2 model, K2.5 underwent continued pretraining on approximately 15 trillion mixed visual and text tokens, resulting in a native multimodal model with state-of-the-art coding and vision capabilities.
A key innovation is the self-directed agent swarm paradigm. For complex tasks, K2.5 can automatically create and orchestrate a swarm of up to 100 sub-agents, executing parallel workflows across up to 1,500 tool calls. This approach reduces execution time by up to 4.5x compared to single-agent setups, without any predefined subagents or workflows.
K2.5 is available across multiple platforms including Kimi.com, the Kimi App, the API, and Kimi Code. The web and app now support four modes: K2.5 Instant, K2.5 Thinking, K2.5 Agent, and K2.5 Agent Swarm (Beta). The Agent Swarm mode is currently in beta on Kimi.com, with free credits for high-tier paid users.
The model excels in coding, particularly front-end development, and can generate complete interfaces from simple conversations. Its vision capabilities allow it to reason over images and video, improving image/video-to-code generation and visual debugging. The company highlights its performance on agentic benchmarks like HLE, BrowseComp, and SWE-Verified at a fraction of the cost.
Kimi K2.5 also introduces significant improvements in office productivity, with benchmarks showing 59.3% and 24.3% improvements over K2 Thinking on the AI Office Benchmark and General Agent Benchmark, respectively. The model can handle tasks such as adding annotations in Word, constructing financial models with Pivot Tables, and writing LaTeX equations in PDFs, scaling to long-form outputs like 10,000-word papers.