The size of BERT's vocabulary depends on the specific implementation and pre-training configuration. BERT models typically use WordPiece tokenization, which splits words into subwords or morphemes. The size of the vocabulary is determined by the number of WordPiece tokens used.

  The original BERT base model, released by Google, has a vocabulary size of 30,000 tokens. This vocabulary is constructed by using a combination of word-level and subword-level tokenization. Words that are not in the vocabulary are further split into subword units.

  It's important to note that the vocabulary size can be different in other BERT models or custom implementations. For example, models fine-tuned for specific tasks may have smaller vocabularies that only cover the relevant domain.

  In summary, the vocabulary size of BERT is typically around 30,000 tokens, but it can vary depending on the specific implementation or configuration.