The performance is good once the tokenizer is loaded, but we spend most of our cycles in the cJSON backend when loading large token mappings.