Term
Kimi K2.5
Kimi K2.5 is an open model from Moonshot AI released in January 2026. It is a native multimodal Mixture-of-Experts model with around 1 trillion parameters (32 billion active), a 256,000-token context and agentic capabilities.
Kimi K2.5 — explained in detail
Kimi K2.5 is an open-weight language and agent model from the Chinese provider Moonshot AI, released on 27 January 2026. The weights are freely available via platforms such as Hugging Face. Kimi is also the name of Moonshot AIs associated chatbot service.
Architecturally, K2.5 is a Mixture-of-Experts (MoE) model with around 1 trillion parameters in total, of which about 32 billion are activated per token. The model has 384 experts (8 per token plus one shared expert), 61 layers and a context window of 256,000 tokens. K2.5 is natively multimodal and understands images and video in addition to text; training involved around 15 trillion mixed text and image tokens.
A key focus lies on agentic capabilities. The model offers an instant and a thinking mode as well as an “Agent Swarm” feature that can coordinate up to 100 specialised sub-agents in parallel. This makes Kimi K2.5 specifically geared towards autonomous, multi-step tasks.
Example / practical relevance
Kimi K2.5 is suited to complex, tool-assisted workflows — such as research, data analysis or software development, where several sub-tasks are handled in parallel. Its native multimodality also allows image and video material to be included directly in such agent workflows. As an open-weight model, it can be run locally or on your own infrastructure.
Distinction
As an open-weight model, Kimi K2.5 sets itself apart from purely proprietary models (GPT, Claude, Gemini). Among open families (DeepSeek, Qwen), it stands out through its native multimodality and Agent Swarm coordination. The version step K2.5 places the model between the K2 line and later versions (such as K2.6).