Kimi K2: Chinese LLM for agents, code, and tool usage

Kimi K2 shifts the conversation about Chinese LLMs from chat towards agents: models that use tools, write code, and operate in long context.

Kimi K2, developed by Moonshot AI, is one of the most interesting Chinese models not because it tries to be another general chatbot, but because it strongly emphasizes agentic work: using tools, coding, solving multi-step tasks, and maintaining long context.

In practice, this direction may be more important than the classic competition for conversational fluency.

Companies are increasingly asking whether a model can write a nice paragraph.

They ask if it can analyze a repository, prepare a patch, call a tool, interpret the result, and return to the plan after an error.

The Kimi-K2-Instruct model card on Hugging Face describes Kimi K2 as a mixture-of-experts model with about one trillion total parameters and 32 billion active parameters.

This is similar economic logic to other large MoE models: large capacity for the entire network, but limited cost for a single response because only a part of the experts is active.

The model card also indicates a context window of 128K and an emphasis on "agentic intelligence," meaning the ability to use tools and perform tasks.