Kimi is the assistant from Chinese lab Moonshot AI, whose open-weight K2 models jumped to the top of global benchmarks for agentic and coding tasks, and whose free assistant made million-token context mainstream in China before most of the world had heard of it.

Kimi built its reputation on extreme long-context: reading hundreds of documents or entire books in one conversation. Its K2 generation, released with open weights, ranks among the strongest models anywhere for tool use, multi-step agentic work, and code, and the K2 Thinking variant runs long chains of reasoning with tool calls. The assistant layers web search, file understanding, and a researcher mode over the models, free on web and mobile, while the API serves the same models at aggressive prices and open weights allow self-hosting.

Who is it for?

Developers wanting top-tier open models for agents and code, researchers processing massive documents, and cost-sensitive teams benchmarking against Western APIs. Hosted-service users should weigh Chinese data jurisdiction; self-hosting sidesteps it.

How much does Kimi cost?

The assistant is free with generous limits. API pricing is usage-based and markedly cheaper than Western frontier equivalents; open weights make self-hosted deployment compute-only.

Our verdict

Kimi K2 is arguably the strongest open-weight line for agentic work, and the free assistant is superb for long documents. For agent builders optimizing cost against capability, it has become impossible to ignore.