My understanding is that transformer model scale quadratically with the sequence length so it makes sense to encode your text into as few tokens as possible.

In any case, I don't understand why this is called 'Token-Free'. There are still tokens, there is just one for each character.