Moonshot AI launches Kimi K2.5 with swarm of 100 parallel agents

Moonshot AI has officially launched Kimi K2.5, the latest open-source multimodal language model, available to users on Kimi.com, the Kimi App, API, and Kimi Code. The release targets developers, enterprises, and AI researchers seeking advanced coding, vision, and agentic computing capabilities.

K2.5 introduces a novel agent swarm mechanism, enabling up to 100 sub-agents to execute 1,500 tool calls in parallel, reducing complex workflow execution time by as much as 4.5 times compared to single-agent models. This feature is currently in beta for high-tier paid users, with broader access through the API and integration into popular IDEs.

Built on continued pretraining over 15 trillion visual and text tokens, Kimi K2.5 achieves state-of-the-art performance in both front-end and multi-language code generation, vision-to-code tasks, and visual debugging. Compared to prior versions like Kimi K2, the new model shows consistent improvements in end-to-end tasks, including:

1. Document creation
2. Spreadsheet modeling
3. Large-scale office work

Benchmarks indicate strong results at lower costs when compared to leading alternatives.

Moonshot AI developed Kimi K2.5 using Parallel-Agent Reinforcement Learning, which allows a trainable orchestrator to dynamically create and coordinate subagents for distributed task execution. This architecture addresses previous bottlenecks and pushes the boundaries of scalable AI, positioning Moonshot AI as a competitive force in open-source AGI research.

Source