I'm having a very hard time finding mention of token-generation speed.