Qwen Image 2.1 rate limits: 120 RPM Turbo, 20 RPM Pro, and batching

Alibaba's Model Studio pages list 120 requests per minute for qwen-image-2.1-turbo and 20 requests per minute for qwen-image-2.1-pro, in every region. At those caps a 1,000-image run takes at least 9 minutes on Turbo (1,000 / 120) and at least 50 minutes on Pro (1,000 / 20), if every request makes one image and none is rejected.

The arithmetic

The figures below are plain division of the vendor's request cap. They assume one image per request and no retries, and the vendor pages do not say whether a request can return more than one image.

Sume's side of a batch

Sume does not list Qwen-Image-2.1. For the Qwen rows it does list, the Image API waits up to 30 seconds and then returns 202 with a job envelope. With mode: "async" the call returns the envelope at once and you read the images from GET /v1/jobs/{id}/result. Slow configurations fall back to 202, so branch on the status code, not the body shape.

The sample submits four async jobs on qwen/qwen-image and prints the job ids.

import os, requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
ids = []
for i in range(4):
    r = requests.post(
        "https://api.sume.com/v1/images",
        headers=H,
        json={"model": "qwen/qwen-image", "prompt": f"poster draft {i}", "mode": "async"},
        timeout=30,
    )
    r.raise_for_status()
    body = r.json()
    job = body.get("data", body)
    ids.append(job.get("job", job)["id"])
print(ids)

Rules of thumb

• Read the cap from the vendor page for the region you call, not from a blog.
• Back off on 429 and do not resubmit a paid job because a local timer fired; store the job id instead.
• Sume's own limits are separate from Alibaba's; check your plan before you size a run.