A language model never reads letters. Before any text reaches it, a tokenizer chops the text into tokens, where a token is one entry in a fixed list called the model's vocabulary. It is the smallest unit the model reads and writes: not a word, not a character, not a syllable, but whatever unit the tokenizer settled on, usually a subword. What the model actually receives is never the characters themselves, but a list of integer IDs, one per token, where each ID is that token's position in the vocabulary.