Moonshot AI has unveiled Kimi K2.5, described as the most powerful open-source model to date. Building on its predecessor Kimi K2, the new model underwent continued pretraining on approximately 15 trillion mixed visual and text tokens, resulting in a native multimodal architecture that excels in both coding and vision tasks.
A key innovation is the self-directed agent swarm paradigm. Kimi K2.5 can autonomously orchestrate up to 100 sub-agents, executing parallel workflows across up to 1,500 tool calls. This reduces execution time by up to 4.5x compared to a single-agent setup, without any predefined sub-agents or workflows. The agent swarm is trained using Parallel-Agent Reinforcement Learning (PARL), which employs staged reward shaping to encourage parallel execution and prevent serial collapse.
Kimi K2.5 is available via Kimi.com, the Kimi App, API, and Kimi Code. The platform now supports 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.
On agentic benchmarks including HLE, BrowseComp, and SWE-Verified, Kimi K2.5 delivers strong performance at a fraction of the cost. It is particularly noted for front-end development, capable of converting simple conversations into complete interactive interfaces with animations. The model also excels in image/video-to-code generation and visual debugging, thanks to massive-scale vision-text joint pre-training.
For real-world software engineering, Kimi K2.5 shows consistent improvements over K2 on the internal Kimi Code Bench. The new Kimi Code product, which is open-sourced, integrates with IDEs like VSCode and Cursor, and supports images and videos as inputs. Additionally, K2.5 Agent handles office productivity tasks such as document creation, spreadsheet modeling, and PDF editing, achieving 59.3% and 24.3% improvements over K2 Thinking on internal benchmarks.