Kimi K3: China's giant AI is here and anyone can try it for free

Beijing-based Moonshot AI has released Kimi K3, its largest language model so far. With 2.8 trillion parameters, the company calls it the biggest AI model whose weights are set to be opened up. The weights are the trained model data itself: with them, anyone can download the AI, run it on their own hardware and adapt it instead of using it only as a cloud service. Trying it does not require a developer account: K3 is now live in the Kimi app for iOS and Android, on kimi.com and in the Kimi Work desktop app. A free account is enough for a first look. Moonshot adjusts the exact usage limits in the app.

Behind Fable 5, ahead of the rest

Moonshot is unusually honest about where its model stands. In the company's own benchmark table, K3 trails Claude Fable 5 and GPT-5.6 Sol but beats Claude Opus 4.8 and GPT-5.5, among others. Early independent numbers point the same way: Artificial Analysis places K3 behind only Fable 5 in its private evaluation, and K3 tops Arena.ai's frontend coding ranking. Most other figures still come from Moonshot itself. On the technical side, K3 offers a 1-million-token context window, handles images natively and, as a mixture-of-experts model, activates 16 of its 896 experts per request. Its predecessor Kimi K2 already made a name for itself with strong benchmark results.

The open weights only arrive on July 27

This matters for anyone who wants to run the model on their own hardware: the open weights are not out yet. Moonshot promises them by July 27. Until then, K3 runs only through the app, the web and Moonshot's own API. There the model costs $3 per million input tokens and $15 per million output tokens, more than any Chinese model before it. On top of that, K3 currently thinks at maximum effort on every request. That makes answers thorough, but slow. Lighter modes are planned for later.

What users should keep in mind

Anyone testing K3 sends their prompts to the servers of a Chinese provider. Sensitive content deserves the same caution as with any cloud chatbot. Moonshot also admits that K3 still trails GPT-5.6 and Fable 5 in user experience. For a free look at the current state of China's AI push, it is still the most interesting model to try right now.