Updated on September 24, 2026
Qwen-Image 2.1 is Alibaba's open-weight image model, covering generation and editing in one checkpoint whose visual component holds just 7 billion parameters. Its signature trick is direct output of transparent RGBA images, with no background-removal pass. The weights are free to download, though the license restricts usage to research, a detail worth reading before any commercial project. It also accepts up to 10 reference images and renders at native 2K.
Qwen-Image 2.1 natively generates RGBA images with a real alpha channel, where most models force a separate background-removal tool into the pipeline. It reads your prompt and decides between a regular picture and a transparent one, edits an existing transparent layer while keeping its empty background, and pulls a subject out of an ordinary photo as an isolated element. You send a sneaker shot, ask for the product alone, and the cutout comes back at 2K.
The rest of the spec sheet fits in a few lines.
| Task | What Qwen-Image 2.1 does |
|---|---|
| Text to image | Native 2K output across several aspect ratios |
| Transparent image | Alpha channel generated directly, editable layers |
| Local editing | Areas marked with circles, masks or brush strokes |
| Multi-reference | Up to 10 input images, group portraits or virtual try-on |
| Subject extraction | An object lifted from a photo as an RGBA layer |
According to Alibaba, on its own benchmark, Qwen-Image 2.1 beats most closed models, Google's Nano Banana 2.0 included, with independent evaluations still pending. Treat that claim with care, although one detail is hard to ignore, since on that same in-house chart every open model listed, FLUX 2 among them, sits below it.
The race between text-to-image models usually runs on tens of billions of parameters, and Alibaba went the other way with a far lighter engine. Against hosted services such as Seedream 4.0 or ChatGPT Images 2.0, the pitch is control rather than raw score. Everything runs on your own machine.
The open weights released live on Hugging Face, GitHub and ModelScope, alongside an online demo and day-one support in ComfyUI and Diffusers. Early reports describe the model running on a solid consumer card like an RTX 3090, and quantized builds trim the memory bill further. Alibaba publishes no official VRAM figure, so plan for a bit of trial and error.
Then comes the awkward chapter. The Qwen Research License covers research and evaluation only, and commercial work requires a separate Qwen license requested by email (anyone used to the Apache 2.0 terms of earlier releases should reread the paperwork). Access inside Qwen Chat and any Alibaba Cloud pricing remain to be confirmed on official channels.
Qwen-Image 2.1 is Alibaba's open-weight image model, combining text-to-image generation and image editing in a single checkpoint. Its visual generation component holds 7 billion parameters, it outputs native RGBA transparency and 2K resolution, takes up to 10 reference images, and its weights download from Hugging Face, GitHub and ModelScope.
No, downloading is free but commercial usage is not covered. The Qwen Research License Agreement limits the model to research and evaluation, and a separate commercial license must be requested from Alibaba by email, with a Built with Qwen notice required in product documentation. Earlier Qwen-Image releases shipped under Apache 2.0, which explains the frequent confusion.
Alibaba publishes no official VRAM figure. Early reports describe the model running on consumer cards such as the RTX 3090 and RTX 4090, quantized and GGUF builds lower the memory footprint further, and the official code can offload part of the model to system RAM on smaller cards. A few test runs will settle your own setup.
Pick based on where you want the model to run. Nano Banana 2.0 is a hosted Google service, while Qwen-Image 2.1 installs on your own hardware and generates native RGBA transparency directly. According to Alibaba, on its own benchmark, it edges ahead of its rival, with independent evaluations still pending.
Verdict: Local logo, sticker and product-cutout work is where this small model replaces two tools at once. Researchers, ComfyUI tinkerers and teams willing to sort out the commercial license will find a very serious testing ground.
Deevid AI Image 
ChatGPT Image 2.5 
Nano Banana 2 
