What is a context window

A model's context window is how much text it can hold in mind at once, measured in tokens — word-pieces, where a thousand tokens is roughly 750 English words. Everything shares the window: your question, any documents you paste in, the conversation so far, and the space the reply needs. It is a working area, not a memory — when a conversation outgrows it, the oldest parts fall out.

Sizes on this site range from a few thousand tokens to over a million. As a rough feel: 128K tokens is a long novel, and a million is a small codebase. The catalogue lists each model's published window, and the model pages translate it into words for you.

Three cautions that the headline number hides:

Served context can be smaller than published. A host may run a 256K model with a 64K window to save memory. The provider table on each model page shows what each host actually serves — the same model, rented from two places, can accept very different amounts.

Big windows cost twice. You pay per token sent, so filling a large window makes every question expensive; and models genuinely get worse at finding things in the middle of a very long context. More is not automatically better.

Running locally, the window costs memory. The model's weights are only part of what fills your graphics memory — the context takes its own share, growing with size. Our fit verdicts show usable context per device for this reason: a model can "fit" with almost no room left to think in.

Where next: What is quantization · When to consider running models locally