image questions with qwen3.5

[afaulconbridge]

I'm trying to identify animals in video clips by sending a number of frames to qwen3.5 with a text prompt containing background information. This works well in the build-in web interface of llama.cpp 👍

However, I have hundreds of clips so I'm scripting against the OpenAI-compatible API of llama.cpp and that fails. My message looks like:

[
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "{text prompt}"
            },
            {
                "type": "image_url",
                "image_url": "data:image/jpeg;base64,{image-in-base64}"
            },
            {
                "type": "image_url",
                "image_url": "data:image/jpeg;base64,{another-image-in-base64}"
            }
        ]
    }
]

The response I get is {'error': {'code': 500, 'message': 'Invalid url value', 'type': 'server_error'}} but if I use the same message against the OpenAI API it works.

Is this a limitation of the llama.cpp OpenAI-compatible API? I couldn't see anything documented that this wouldn't work, but it is a bit of an edge case. Searching is a messy experience though as it keeps returning image-generation related documentation which isn't appropriate here.