Qwen-Image 2.1 is 7B: self-host it or call hosted qwen-image on Sume

Qwen-Image 2.1 is a 7-billion-parameter visual generator, and its Hugging Face model card (read 2026-10-05) names the Qwen Research License Agreement. Sume does not host 2.1. Sume's catalog has qwen/qwen-image and qwen/qwen-image-max, which are other rows, listed at $0.02 and $0.075 per image before margin. Read the licence before you plan any commercial use of the weights.

What the model card says

Facts from the card, read 2026-10-05:

Hosted cost on Sume

Sume bills provider list x 1.25. These are the Sume rows, not Qwen-Image 2.1, and the arithmetic is before cent rounding. The live endpoints pricing line is authoritative.

Self-host or call an API

Self-hosting means you provide a GPU, handle the Diffusers install, and carry the licence question. We did not measure memory use or speed, and the model card text we read does not give a verified VRAM figure, so test on your own hardware before you size anything. A hosted call costs by the image and needs nothing local, but you get the hosted model, not 2.1.

A rough break-even: if the GPU you would rent costs $X per hour, divide by your images per hour to get a per-image figure, then compare it with $0.025. Include idle time, which is where self-hosting usually loses.

A hosted call

This calls the hosted row on Sume. Because 2.1 is not on Sume, results will differ from the 2.1 weights.

import asyncio, os, httpx

async def main():
    body = {"model": "qwen/qwen-image", "prompt": "a bakery storefront at sunrise, poster style", "aspect_ratio": "3:4"}
    headers = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
    async with httpx.AsyncClient(timeout=60) as c:
        r = await c.post("https://api.sume.com/v1/images", headers=headers, json=body)
    print(r.status_code, r.json().get("usage"))

asyncio.run(main())

What we did not check

We did not run the model, so this post makes no claims about speed, memory or output quality. Hosted and self-hosted are two different products here: the hosted rows are not Qwen-Image 2.1, so a hosted result tells you nothing certain about 2.1.