You type a question and Kimi writes back in a few seconds. What happens in between? This page explains it in plain words: how Kimi reads your message, how it "thinks", how it uses tools like web search, and why it is sometimes wrong. No technical background needed.
The whole idea in one picture: read the pieces, work out what comes next, write it down.
Kimi turns your words, files and pictures into small pieces it can work with.
A very large trained model works out, piece by piece, what a good reply looks like.
For bigger jobs it can search the web, use tools and check its own work.
Your phone or computer sends it to Moonshot's servers. The heavy work happens there, not on your device.
The text is cut into small pieces and turned into numbers the model can use.
It puts together your new message, the earlier chat and any files you attached.
For fresh facts or bigger jobs, it may search the web or use another tool first.
On harder questions, the model works through the problem before it answers.
The answer is produced one token at a time. That is why you see it appear gradually.
Kimi does not read whole sentences the way you do. It reads tokens: short words, parts of longer words and punctuation marks. Limits and API prices are counted in tokens, so the idea is worth knowing.
This is a simple illustration that runs in your browser. It is not Kimi's real tokenizer, and nothing you type is sent anywhere. Real counts will differ, especially for languages other than English.
The context window is everything Kimi can keep in view at one time: your messages, its replies and your files. Think of it as the size of its desk. Anything that does not fit on the desk cannot be used for the answer.
Uses about 650 tokens a page, which is a rough estimate. Model limits are the API figures: 262,144 tokens and 1,048,576 tokens. In the Kimi app, the limit you get can depend on your plan. Leave room for Kimi's reply too.
A very long chat can also fill the window. If Kimi starts to forget early details, begin a new chat and paste in the key points. More: summarizing long PDFs.
Kimi is a large language model. Before anyone could use it, it was trained on a huge amount of text. During training it practised one thing again and again: guessing the next piece of text. Each wrong guess nudged its internal settings a little.
Those internal settings are called parameters. Kimi K3 has about 2.8 trillion, the most of any openly released model.
The model does not store documents to look up. It holds patterns of language and knowledge learned from them.
Training stops on a certain date. For anything newer, Kimi needs web search or a file from you.
Moonshot has not published a full list of what its models were trained on. Model details: Kimi K3 · choose a Kimi model
For simple questions, a quick reply is fine. For hard ones, Kimi can first write out its reasoning to itself, step by step, and only then give the answer. This usually improves results on maths, planning and code. It also takes longer and uses more tokens.
| Model | Thinking |
|---|---|
| Kimi K3 | Always on. Developers can set the effort to low, high or max |
| Kimi K2.7 Code | Always on |
| Kimi K2.6 | Can be switched on or off |
From Moonshot's API documentation, as collected for this site. The modes offered in the Kimi app may be named differently.
By itself, a language model can only write text. Kimi becomes far more useful because the app connects the model to tools. The model asks for a tool, the tool does its job, and the result is handed back for the model to use.
Finds current pages, so answers are not limited to what the model learned.
Turns your PDFs, documents and images into text the model can read.
Produces real documents, slides, spreadsheets and websites.
Pulls in figures such as market and economic data.
See them in use: Deep Research · Kimi Docs · Professional Data · all features
In a normal chat, Kimi answers once and stops. As an agent, it keeps going. It makes a plan, takes a step with a tool, looks at the result, fixes what went wrong, and repeats until the task is finished.
A swarm goes one step further. A lead agent splits a big job into parts, hands them to many helper agents working at the same time, then joins their results. Moonshot's figures are up to 300 helper agents and 4,000 steps.
More: Agent Mode · Agent Swarm · Kimi Claw
The model runs on Moonshot's servers. Your device sends the request and shows the reply, so you need an internet connection.
Developers send requests to the same models from their own software and pay per token.
Kimi's model weights are open. Cloud providers and expert teams can run it themselves, though Kimi K3 needs very large hardware.
Because your messages are processed on a server, read Kimi's own privacy terms before sharing anything sensitive. More: Kimi API · download Kimi
Kimi writes what is likely, which is not always what is true. Knowing why helps you get better answers.
More: how to use Kimi AI · Kimi AI review · common fixes