Qwen-Image-2.1-Turbo runs 8 steps; what does Sume's Qwen row expose?

Sume's qwen/qwen-image row does not expose denoising steps, CFG or a seed. Its catalog lists prompt, aspect_ratio, n from 1 to 4, input_references from 0 to 10, and output_format of png, jpeg or webp, plus custom pixels through image_size. The Qwen-Image-2.1-Turbo model card is about an 8-step checkpoint you run yourself, and Sume's docs do not say which Qwen version its row serves.

What does the Turbo model card describe?

The card describes a 7B model that runs in 8 steps, supports RGBA transparency, edits through reference images with masking, and ships under the Qwen Research License. Those are settings and terms for someone downloading weights. A hosted API row is a different contract: the provider fixes the steps and Sume publishes the fields it accepts.

What happens if I send steps or seed to Sume?

Sume does not drop unknown knobs silently. Its docs say that a request using a parameter the selected model does not list returns 400 unsupported_parameter. seed and output_compression are in the schema, but no model advertises them in v1, so they return the same error everywhere. Fields the schema does not name at all, such as num_inference_steps, are not part of the contract.

Does a faster checkpoint change the price on Sume?

Not through any field you can set. Sume meters image models per image: the cost in a response is the billed USD amount, and the token counts in usage are always 0 in v1. The qwen/qwen-image row is a flat $0.025 whatever the provider does internally, and a failed or cancelled generation is not billed. Fewer steps is a speed story for people running weights; on a per-image price, only the row you pick changes the bill.

How do I see exactly what the row accepts?

Ask the endpoint record, which is the definitive list per Sume's docs. The script prints the parameter names and the ratio list.

import os
import requests

url = "https://api.sume.com/v1/images/models/qwen/qwen-image/endpoints"
r = requests.get(
    url,
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    timeout=30,
)
r.raise_for_status()
ep = r.json()["endpoints"][0]
params = ep["supported_parameters"]
print(sorted(params))
print(params["aspect_ratio"]["values"])
print(ep["pricing"])

When should I self-host instead?

If you need exact steps, a fixed seed or native RGBA from the model, run the weights yourself and read the license first. If you need a pay-per-image call with one set of fields, use the hosted row, and for a transparent PNG switch to openai/gpt-image-2.5 with background: transparent. Slow calls past 30 seconds come back as 202; see Jobs and results.