The context window refers to the upper limit of tokens a model can process simultaneously in a single inference; 128K is roughly equivalent to a medium-length book, and 1 million tokens is roughly equivalent to thousands of pages of documents or a small-to-medium-sized codebase.