Give a text-only dsh model eyes: pasted images recognized into text via an OpenAI-compatible vision endpoint.