Maximum sequence length a language model can attend to at once, measured in tokens.