It is a well-known fact that different model families can use different tokenizers. However, there has been limited analysis on how the process of “tokenization” itself varies across these tokenizers.