Is there any details about hardware requirements for a sensible tokens per second for each size of these models?