The fixed set of tokens a model can read and write, typically tens to hundreds of thousands of entries.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
A model's vocabulary is decided before training and never changes. Every input must be expressible in it and every output is drawn from it, which is why an unusual word arrives as several fragments rather than as itself.
Vocabulary size is a real design tradeoff: larger vocabularies mean shorter sequences and cheaper attention, but a bigger embedding table and output projection, and less data behind each rare entry.