Kimi K2.5: Open-Source Multimodal Agentic Model with Agent Swarm
A new open-source model called Kimi K2.5 has been unveiled by researchers to enhance general agentic intelligence. This model is designed to optimize text and visual processing through innovative methods, including joint text-vision pre-training and reinforcement learning. It features an Agent Swarm system, enabling collaboration among multiple agents to efficiently tackle complex tasks by breaking them into smaller, manageable components. Testing has demonstrated that Kimi K2.5 significantly outperforms in areas like coding and reasoning, while Agent Swarm can reduce processing latency by up to 4.5 times compared to traditional single-agent approaches. A post-trained model checkpoint is now available for further investigation.
Key facts
- Kimi K2.5 is an open-source multimodal agentic model.
- It emphasizes joint optimization of text and vision.
- Techniques include joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning.
- Agent Swarm is a self-directed parallel agent orchestration framework.
- Agent Swarm decomposes complex tasks into heterogeneous sub-problems and executes them concurrently.
- Kimi K2.5 achieves state-of-the-art results in coding, vision, reasoning, and agentic tasks.
- Agent Swarm reduces latency by up to 4.5 times over single-agent baselines.
- The post-trained model checkpoint is released.
Entities
—