Confused as to why you wouldn’t integrate a local vlm if you want a local llm as the backbone. Plenty of 8b - 30b vlms out there that are visually competent.