21 September 2026
Qwen released Qwen-Image-2.1 on 20 September 2026, an image generation and editing model with 7 billion parameters in its generator. On Qwen’s own quality ranking it scores 60.28, ahead of every other model with downloadable weights. The catch is the licence: unlike every earlier open Qwen image model, this one is for research only.
What can Qwen-Image-2.1 do?
One model does both jobs. Write a prompt and you get an image. Hand it an image and an instruction and you get an edit. Earlier Qwen releases split those across separate models. The generator has 32 Single-Stream DiT layers, reads prompts through a Qwen3-VL 8B text encoder, and renders natively at 2K, up to 2752 by 2752 pixels.
The new capability is transparency. Qwen-Image-2.1 outputs RGBA images, meaning images with a see-through background, straight from a prompt. It can edit a transparent layer without flattening it, and it can lift a subject out of an ordinary photo as a transparent cut-out. Qwen shipped that as a separate model, Qwen-Image-Layered, in December 2025. It now sits inside the main model.
Editing accepts up to 10 reference images at once. Qwen’s examples merge six portraits into one group photo and ten furniture shots into a single room. You can point at a region three ways: draw coloured circles on the image, paint over the area, or pass the original plus a separate mask. The mask route is the one to use when you cannot afford to paint over the original pixels.
Qwen-Image-2.1 benchmarks against GPT Image 2.5 Sunburst and Nano Banana 2.0
Qwen ranks 30 image models on a benchmark it calls Qwen-Image-Bench. Here is where the new model lands against the leaders, its nearest open rival and its own predecessor.
| Model | Qwen-Image-Bench total score | Parameters as Qwen lists them |
|---|---|---|
| GPT Image 2.5 Sunburst (OpenAI) | 67.01 | Not disclosed |
| Qwen Image 3 Pro (Qwen, closed) | 62.36 | Not disclosed |
| Qwen-Image-2.1 (Qwen, open weights) | 60.28 | 7B |
| Nano Banana 2.0 (Google) | 59.82 | Not disclosed |
| FLUX 2 Max (Black Forest Labs) | 55.33 | 32B |
| Qwen-Image-2512 (the previous Qwen open model) | 52.06 | 20B |
These are Qwen’s own numbers on a benchmark Qwen built and has not released. The announcement publishes the comparison as a chart image, with no table, no per-category scores and no statement of how the rival models were scored or which prompts were used. Treat the ordering as a claim rather than a measurement. It is still useful for one thing: it tells you which models Qwen considers the competition.
How to run Qwen-Image-2.1, and what the licence allows
- Licence: Qwen Research License Agreement, non-commercial use only unless you get a separate commercial licence
- Download: Hugging Face and ModelScope, as Qwen/Qwen-Image-2.1
- Runtime: Diffusers, ComfyUI, vLLM-Omni, SGLang and LightX2V all shipped support on day one
- Defaults: the model card example runs 40 steps at 2048 by 2048 pixels
- Extras: two prompt-rewriting models, PE-T2I and PE-I2I, built on Qwen3.5-VL 9B, turn a prompt in any language into English
- Memory: the model card offers CPU offload but publishes no VRAM figure
The licence is the change that matters. Qwen-Image, released in August 2025, and Qwen-Image-2512 after it were both Apache 2.0, which lets you sell what you build. Qwen-Image-2.1 is not. The announcement calls the release open-source and the weights really are downloadable, but the licence file says you may not use the model for any commercial purpose without a separate agreement. If you generate product shots for a webshop, that is the difference between shipping and not shipping.
What Qwen is not saying about Qwen-Image-2.1
Read the ranking again and you find Qwen Image 3 Pro at 62.36, above the model Qwen just released. That is Qwen’s own closed model, and the announcement text never mentions it. The open release is deliberately the second-best image model Qwen has.
Efficiency is the headline of the post, yet the post contains no timing at all. No seconds per image, no VRAM figure, no throughput on a named card, no API price. The architecture section explains why the model should be fast, with a mixed-granularity attention scheme that caches your reference images once instead of re-reading them at every step. It never says how fast that turns out to be. For a release whose whole pitch is cost, that is the number missing from the page.
What this means
Worth testing now if you make product or marketing images and can run your own GPU, and worth testing this week if transparent assets are part of the job. A three-person design studio cutting products out for a webshop today pays for a background remover, a generator and an editor. Qwen-Image-2.1 covers all three at a size that fits one workstation instead of a cluster. The 60.28 against 59.82 for Google’s Nano Banana 2.0 also says the quality gap to the closed leaders is now small enough to argue about.
Do not put it in a paying product yet. The research licence means a commercial deployment needs a separate agreement with Alibaba, and Qwen has published neither the terms nor a price, so this is a procurement conversation rather than a download. The pattern is worth noticing too. Qwen kept its image models on Apache 2.0 through 2025 and has now put the good one behind a research licence while a closed Qwen Image 3 Pro sits above it in its own ranking. If your plan assumed Qwen weights would stay permissive, this is the release that should prompt you to check. What I would test first is transparency on your own product photography, because it is the one thing nothing else open does in a single model, and Qwen shows it only on hand-picked examples.
For more information, visit the official announcement of Qwen-Image-2.1 on the Qwen blog.
