Send images to the API
Vision on ElephantPool: which models accept images, how to send one with the OpenAI SDK, the limits, and why image URLs are refused by design.
Models tagged vision in the catalogue accept images alongside text — the standard
OpenAI multimodal format, so your SDK already knows how.
Which models
Check GET /v1/models or the models page for the
vision capability. A model that cannot see refuses images with a clear 400
rather than silently ignoring your picture and answering about nothing — you should never pay for
an answer to a question the model never saw.
Sending one
client.chat.completions.create(
model="qwen3.8-27b",
messages=[{
"role": "user",
"content": [
{"type": "text", "text": "What is in this picture?"},
{"type": "image_url",
"image_url": {"url": f"data:image/png;base64,{b64}"}},
],
}],
)Inline the bytes — links are refused, and that protects you
Only data: URLs are accepted; an https link is rejected. The reason is
the architecture: your request runs on a community machine, and following a link would make that
machine fetch an arbitrary address on your behalf — and make your prompt's context
dependent on whatever that server returned at that moment. Inlining keeps the job self-contained:
what you sent is exactly what the model saw.
Limits and cost
- Up to 6 MB of base64 per image (roughly a 4.5 MB file) — over the limit the error says so explicitly.
- Images count through the model's normal token pricing; there is no separate image fee.
- Like every prompt: never written to disk, never trained on. The image exists in memory for the duration of the job.
See also: the model catalogue · download mahout · how the whole system works · the blog