Has anyone seen any lightweight models that could pre-optimize prompts for reduced token usage?

interesting question, I haven’t actually fine-tuning small models for this particular task and getting into inference with larger model would be really interesting? 😱