Context Window
The context window is how much text a model can consider at once, measured in tokens — and holding that text costs memory through the KV cache.
Models read and write in tokens — word fragments, roughly 750 English words per 1,000 tokens. The context window is the total the model can hold in view at once: your instructions, the conversation so far, any pasted documents, and the answer being written.
Publishers quote ever larger windows, and the numbers are real — but on your machine the window is not free. Every token held in view occupies KV-cache memory, so a huge advertised window and a modest graphics card meet in the middle: the model SUPPORTS a million tokens, your memory decides how many you can actually use.
That is why any honest fit statement names its context assumption. The same model on the same card can be a comfortable everyday fit at short contexts and an impossibility with a whole book loaded.
Where you see it on the radar
Every fit estimate on the radar states the context length it was computed at — change the assumption and the verdict honestly changes with it.
See fits with their context stated →Related terms