Generate with the Qwen image models using Model Studio's native messages shape: the prompt and any reference images all live in input.messages[].content[]. Existing Model Studio clients only need to point base_url at Amux.

This one is synchronous: the connection stays open and the image comes back on it. Measured latency is around 48s at 1024x1024 and 68s at 2K -- that is not us being slow, the vendor's synchronous endpoint takes that long. Use the asynchronous endpoint below, or the unified entry point, if you would rather not hold a connection.

Billing is per image, per resolution tier

The tier (1K / 2K) is decided upstream from the pixel area and reported back in usage.output_image_type. We reserve against an upper bound at submit time and settle on the tier the upstream actually charged.

Reference images

Put 1-3 {"image": "..."} items in content[]; both public https URLs and data:image/...;base64,... are accepted. Image-to-image also bills per input image.

{
  "model": "alibaba/qwen-image-3.0",
  "input": {
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "text": "A red paper boat drifting down a rain gutter"
          }
        ]
      }
    ]
  },
  "parameters": {
    "size": "1024*1024"
  }
}