Context window
The maximum number of tokens a model can attend to at once, including both your input and its output.
Two numbers get called the context window and they frequently disagree. A model's configuration declares what it was trained to handle; the file you download often declares something smaller, and the runtime defaults to smaller still. We publish both where they differ, because quoting either one unqualified is a wrong claim.
The practical constraint is rarely the model's limit. It is memory: the KV cache is allocated for the full context you request, up front, so asking for 128K costs you the full allocation immediately even if your prompts are 4K.