Definition
Context window is a context window is the maximum amount of text, measured in tokens, that an AI model can consider at once, including your prompt, any documents, the chat history and its reply.
A context window is the model's working memory for a single conversation. Everything the model can see when it writes its next word has to fit inside it: your instructions, the files you pasted, the earlier back-and-forth, and the answer it is writing right now.
The size is counted in tokens, not pages or words. A window of 200,000 tokens holds very roughly 150,000 words of English, though the exact figure depends on the text.
The model does not remember you between chats. Each time you send a message, the app sends the model the whole conversation again, trimmed to fit the window. When a chat grows past the limit, something has to go. Usually the oldest messages are dropped or summarized, and the model quietly loses track of them.
A bigger window means you can paste a whole contract, a long codebase or a stack of research notes and ask questions across all of it, instead of feeding it in pieces.
A large window is a ceiling, not a guarantee of quality. Three things to know:
Imagine a desk. The context window is the size of the desk. A bigger desk lets you spread out more papers at once, but a cluttered big desk is still harder to work at than a tidy small one.
The window is not what the model learned. Training data shapes what the model knows in general. The context window only holds what you put in front of it right now. For more control over what goes in, some teams use retrieval, which fetches only the relevant passages for each question.
The window limits and prices every call during inference. Models with a long window, including some built as a mixture of experts, still count everything in tokens.
Check the window size before you paste in a big document, and remember the model's reply has to fit too.
The app either refuses the request or drops the oldest content, so the model can no longer see or use what was cut.
No. It lets you include more, but long inputs cost more and models can use buried details less reliably. Send only what the task needs.
Mistral opened a public preview of Mistral Large 4 on Tuesday, a mixture-of-experts model with about 1 trillion parameters and 49 billion active, and promised weights by the end of October. The API is a preview. The weights are the launch.