Although the underlying model is unknown. If it expose input token count, the tokenizer may be probable though. Most tokenizer segemnts wildly different in CJK inputs. It can probably be used to fingerprint the tokenizer based on token count if it is using existing tokenizer.