Moonshot has released Kimi K2.5, an open-source native multimodal agentic model built through continual pretraining on approximately 15 trillion mixed visual and text tokens atop Kimi-K2-Base.
- The model integrates vision and language understanding with advanced agentic capabilities, including instant and thinking modes.
- It supports coding from visual specifications and a self-directed "Agent Swarm" execution scheme for complex tasks.
- A recent update on 2026.1.29 removed the default system prompt to prevent confusion and corrected a token in the chat template.
- The model is available via libraries like Transformers, inference providers such as HuggingChat, and local apps using vLLM or SGLang.
The release provides developers with access to a powerful multimodal foundation for agentic workflows and visual reasoning tasks.