A tokenizer is a deterministic string-matching program, it's not made out of weights in the same sense as a neural network itself.
But it could be. It's just less efficient.
But it could be. It's just less efficient.