[Official FAQ] Corrections and follow-up questions welcome.

VL models accept images plus text. They add a vision tower, so memory and dependency needs usually exceed text-only chat at a similar “B” label.

Reality checks:

• Confirm your runtime (Ollama, llama.cpp, vLLM, etc.) supports that VL architecture.
• Resolution and number of images drive VRAM—start with one downscaled image.
• OCR-like screenshots and charts still hallucinate; verify critical facts.

There is no single VRAM number—use vendor notes and your own trial loads.