Kimi
Very long contexts and agentic reasoning from China.
Olya's profile
- What sets it apart
- Kimi is the assistant from Moonshot AI, a Chinese lab, and its signature is very long context: the flagship K3 reaches a million tokens, the others sit around 256K. The other distinguishing trait is open weights (a modified MIT license, on Hugging Face): you can use it through the free app or the API, but also download it and run it yourself, which Western rivals of the same tier don't allow.
- Strengths
- It's built around agentic tool use: the "Thinking" variant chains hundreds of tool calls on its own and speaks the Model Context Protocol natively. API pricing is aggressive, roughly $0.95 per million input tokens against much higher figures from competitors, and the K2.7 Code version is tuned specifically to work across entire code repositories.
- When to use it
- It's worth it when you need a model to reason over huge documents or codebases without chopping them up, or when you want an agent that runs long chains of actions with minimal supervision. It's also a sensible pick if budget matters or you want open weights to host on your own infrastructure.
- When to avoid it
- If you handle sensitive data and it bothers you that the app routes through Chinese servers, consider self-hosting or look elsewhere. For short everyday tasks the million-token context is wasted, and on Italian and European languages performance stays a notch below the best Western models; for images or voice it isn't the right home.