Can I use Qwen-Image-2.1 commercially?

No. Qwen-Image-2.1 is published under the Qwen Research License Agreement, which grants rights for non-commercial purposes only; The Decoder reports that business users must apply to Qwen for a separate license. That is a change from the earlier Qwen-Image line, whose repository states it is licensed under Apache 2.0.

What hardware does Qwen-Image-2.1 need?

Qwen-Image-2.1 keeps its visual generation component to 7B parameters, and The Decoder reports the model runs on a capable consumer GPU such as an RTX 3090. The model card lists the software side: torch 2.4.0 or newer, transformers 5.17 or newer, diffusers installed from GitHub, plus accelerate and pillow.

How does Qwen-Image-2.1 compare to the previous Qwen image models?

Qwen-Image-2.1 follows Qwen-Image-2.0, which the QwenLM/Qwen-Image README dates to 10 February 2026. The visible differences are the license (research-only instead of Apache 2.0), native RGBA output, and editing from up to 10 reference images. Qwen has not published a score table comparing the two.

Are there published benchmark scores for Qwen-Image-2.1?

Not yet. Qwen claims Qwen-Image-2.1 beats most closed models on Qwen's own benchmark, but the Hugging Face model card carries no benchmark table, and The Decoder notes that independent benchmarks are still pending. Treat the claim as the maker's, not a measured result.

What edit controls and output sizes does Qwen-Image-2.1 support?

Qwen-Image-2.1 steers local edits three ways: you can circle a region, paint an annotation on it, or pass a separate mask. The model card lists seven aspect ratios — 1:1, 4:3, 3:4, 3:2, 2:3, 16:9 and 9:16 — and its quickstart generates at 2048x2048 in 40 inference steps.

Alibaba (Qwen) · 2026-09-20 · major

Qwen-Image-2.1 — a 7B open image model on a research-only license

Qwen-Image-2.1 is Alibaba's new open-weight image model. The 7B visual generator makes and edits pictures, outputs transparent RGBA, and takes up to 10 reference images. It ships under a research license that bars commercial use.

Hugging Face model card banner for Qwen-Image-2.1

Alibaba's Qwen team ships a 7B image generator that edits from ten reference images, under a non-commercial license.

Quick facts

Maker	Alibaba (Qwen)
Size	7B visual generation component
Architecture	32 Single-Stream DiT layers
License	Qwen Research License (non-commercial)
Reference images	Up to 10 per edit
Availability	Hugging Face, ModelScope, demo Space
Requirements	torch >= 2.4.0, transformers >= 5.17, diffusers

What is it?

Native transparency is the headline change in Qwen-Image-2.1: it generates and edits RGBA images directly, so an object or a line of text can sit on a transparent layer instead of a painted-in background. One model covers both jobs — text-to-image generation and image editing — and a single edit can draw on up to 10 reference images, which Qwen points at group portraits, virtual try-ons and room designs.

How does it work?

The visual generation component is 7B parameters across 32 Single-Stream DiT layers, with mixed-granularity attention and prefix KV cache reuse to keep it cheap to run. Edits are steered locally rather than re-described: circle a region, paint on it, or hand the model a separate mask. Qwen ships the model as a diffusers pipeline called QwenImage21Pipeline, loaded from the model id Qwen/Qwen-Image-2.1.

Why does it matter?

The license is the catch. Qwen-Image-2.1 arrives under the Qwen Research License Agreement, which is non-commercial only, while the earlier Qwen-Image repository is Apache 2.0 — so a team that built a product on the open weights cannot simply upgrade without asking Qwen for separate terms. For researchers and hobbyists the trade is a small model that The Decoder reports runs on a consumer card like an RTX 3090, with transparency and multi-reference editing that usually need a hosted API.

Who is it for?

researchers and hobbyists generating or editing images locally

Frequently asked questions

Can I use Qwen-Image-2.1 commercially?: No. Qwen-Image-2.1 is published under the Qwen Research License Agreement, which grants rights for non-commercial purposes only; The Decoder reports that business users must apply to Qwen for a separate license. That is a change from the earlier Qwen-Image line, whose repository states it is licensed under Apache 2.0.
What hardware does Qwen-Image-2.1 need?: Qwen-Image-2.1 keeps its visual generation component to 7B parameters, and The Decoder reports the model runs on a capable consumer GPU such as an RTX 3090. The model card lists the software side: torch 2.4.0 or newer, transformers 5.17 or newer, diffusers installed from GitHub, plus accelerate and pillow.
How does Qwen-Image-2.1 compare to the previous Qwen image models?: Qwen-Image-2.1 follows Qwen-Image-2.0, which the QwenLM/Qwen-Image README dates to 10 February 2026. The visible differences are the license (research-only instead of Apache 2.0), native RGBA output, and editing from up to 10 reference images. Qwen has not published a score table comparing the two.
Are there published benchmark scores for Qwen-Image-2.1?: Not yet. Qwen claims Qwen-Image-2.1 beats most closed models on Qwen's own benchmark, but the Hugging Face model card carries no benchmark table, and The Decoder notes that independent benchmarks are still pending. Treat the claim as the maker's, not a measured result.
What edit controls and output sizes does Qwen-Image-2.1 support?: Qwen-Image-2.1 steers local edits three ways: you can circle a region, paint an annotation on it, or pass a separate mask. The model card lists seven aspect ratios — 1:1, 4:3, 3:4, 3:2, 2:3, 16:9 and 9:16 — and its quickstart generates at 2048x2048 in 40 inference steps.

Try it

Load Qwen/Qwen-Image-2.1 with the diffusers QwenImage21Pipeline, or open the demo Space on Hugging Face.