The tokenizer is the algorithm that splits raw text into tokens for a model's vocabulary — BPE, WordPiece, and SentencePiece are common approaches, and the choice affects what the model sees. A tokenizer trained on English may fragment non-English text into many small pieces, increasing cost and reducing quality for those languages. Vocabulary size is a key trade-off: larger vocabularies mean fewer tokens per text but more parameters in the embedding layer. The tokenizer is often overlooked but directly impacts multilingual performance, code handling, and prompt efficiency.