See sibling answer but essentially the effectiveness of cache is not diminished by having a separate one per model (relative to the win of doing more turns and generation with a cheaper model).

edit: updated the answer above to be more qualitative instead