Comment by embedding-shape

Right, but that doesn't answer why you'd need a fast 7b LLM rather than a slightly less fast 14b LLM.