Has anyone used small agentic models for domain-specific work like medical illustration?

I'm a medical illustrator, and I've been curious about these smaller agentic models (like the 3B parameter ones I see mentioned) for practical tasks – e.g., generating anatomical diagrams or parsing surgeon feedback. But I'm skeptical they can handle the nuance. Anyone here tried them for a specialized, non-generalist field? Id love to hear how it actually went.