Vision (image-to-text) runs on the text mode: submit mode=text, a vision model, your question as the prompt, and the image as init_image.
POST /v1/generate\n{"mode":"text","model":"qwen2.5-vl-7b",\n "prompt":"Describe this image, including any text.",\n "init_image":"https://example.com/photo.jpg"}Live vision models: gpt-4o (frontier), and the self-hosted qwen2.5-vl-7b, llava-1.6, llama3.2-vision-11b, gemma3-4b. The self-hosted ones run on our own GPUs and cost less in FLOP per call. Poll the job id until done; the result is the text answer.