Qwen Image 2.1 Turbo at $0.016 an image vs Qwen Image on Sume

Qwen Image 2.1 Turbo can be called as a hosted API since October 9, 2026, at $0.016 per image in Alibaba Cloud's Singapore region, and Sume does not list it. What Sume lists is a qwen-image row (and a text-only qwen-image-max row), and the qwen-image row works out to $0.025 an image before rounding, so for the same number of pictures the direct Alibaba route is cheaper on list price.

This post puts the vendor's numbers next to Sume's catalog so you can decide whether a different account is worth saving about nine dollars per thousand images. Every Qwen fact below comes from the vendor pages read on 2026-10-10: the Qwen-Image-2.1 repository, the Turbo model card, its license, and Alibaba's Model Studio pricing page.

What Qwen shipped on October 9

The repository's news list has two dated entries: the Qwen-Image-2.1 release on 2026-09-20, and on 2026-10-09 the Turbo variant plus Pro and Turbo APIs on Alibaba Cloud. The base model samples with 40 denoising steps by default; Turbo is tuned for 8. The visual generator has 7 billion parameters, native output is 2K (2048x2048), and the repository says a composition can take up to 10 reference images.

The weights are a separate matter from the API. The Hugging Face card says the model is not deployed by any inference provider, and the weights ship under the Qwen Research License, which allows non-commercial use and requires a separate commercial license from Qwen for anything else. The paid Model Studio API is the vendor's hosted route for commercial work. Neither page states a wall-clock speedup, so treat the 8-step figure as a recipe change, not as a measured latency.

Alibaba's per-image prices, side by side

The pricing page lists qwen-image-2.1-turbo by region. Singapore bills $0.016 per image with a 10-image free quota. China (Beijing), Hong Kong, Germany (Frankfurt), US (Virginia) and Japan (Tokyo) show $0.014133 per image and no free quota. The page also lists qwen-image-2.1-pro at $0.04 per image in Singapore and $0.035333 elsewhere, and it says only successfully generated images are charged.

What Sume lists instead

Sume's image catalog is a fixed list of ids in apps/api/src/schemas.ts; it contains qwen-image and qwen-image-max and no 2.1 Turbo or Pro id. Do not assume the qwen-image row is the 2.1 generation: the catalog names it only as Qwen Image, and the repo's price table keys it to the fal endpoint fal-ai/qwen-image. If a request names an id outside the list, the Image API rejects it; check GET /v1/images/models for the live list before you build on a name.

The two Sume rows differ in what they accept. qwen-image takes reference images for edits, and qwen-image-max is text-to-image only. Both are billed from the catalog formula list x 1.25, rounded up to a whole cent. That is why the table says before rounding: the cent rounding happens on the billed amount, not in the unit rate.

The arithmetic for a real batch

For 1,000 images at the unit rates above, Turbo in Singapore comes to $16.00, Turbo in the other listed regions to $14.13, and Sume's qwen-image to $25.00. The gap against Singapore is $9.00 per thousand, or $90 per ten thousand ($160 against $250). Against the other regions it is $10.87 per thousand.

What the gap buys on the Sume side: no separate Alibaba account or region choice, one wallet and key shared with Sume's video and audio endpoints, and the same job envelope for every call. A request that outlasts the 30-second image wait returns 202 with a job to poll, as described in Jobs and results. What it does not buy is the Turbo model itself.

Which route to pick

Prices on both sides change; re-read the Alibaba page and Sume's catalog before you commit a budget.

• Choose Alibaba's API directly if the target is Qwen Image 2.1 Turbo specifically, you will generate thousands of images, and a second vendor account is acceptable.
• Choose Sume's qwen-image if you already send other media jobs through Sume, want one bill, and can accept the earlier Qwen generation at $0.025.
• Do not download the Turbo weights for a commercial product without the separate license the Qwen Research License calls for.
• Before switching either way, run the same ten prompts through both and compare, since this post compares prices, not quality.