The Qwen image generator is really two models from Alibaba's Qwen team. Qwen-Image-3.0 is the hosted flagship that you use in Qwen Chat or pay for per picture through Alibaba's API. Qwen-Image-2.1, released on 20 September 2026, is the one making headlines. It is a free download with a 7B-parameter image generator that makes transparent PNGs and edits photos with up to 10 reference images.
The catch is in the fine print. The free 2.1 weights ship under a licence that allows "research or evaluation purposes only". The search results are crowded with sites Alibaba does not run. And there is no official way to run the model on a Mac. This guide sorts out which Qwen image model is which, what each costs and where the real one lives. It then covers how Qwen scores against Nano Banana and GPT Image, and what a Mac needs to run it.
The Key Takeaways
- Two models: Qwen-Image-3.0 is hosted only, and Qwen-Image-2.1 (20 September 2026) is the open download with a 7B generator.
- Price: the Qwen Studio app is listed as free on the App Store, and the API charges $0.03 to $0.075 per image depending on model and resolution.
- Licence catch: the 2.1 weights are free to download but limited to research or evaluation, a change from the Apache 2.0 terms of earlier Qwen-Image releases.
- Quality: on Qwen's own benchmark, 2.1 scores 60.28, just ahead of Nano Banana 2.0 (59.82) and well behind GPT Image 2.5 (67.01).
- Mac: Qwen publishes no Apple silicon support. Community ports work, at roughly 2.5 minutes per 1024-pixel image on a 36GB M5 Max.
What Is the Qwen Image Generator?
The Qwen image generator is Alibaba's family of text-to-image and image-editing models, built by the same Qwen team behind the Qwen 3.8 chat models. The family started with the original Qwen-Image, a 20-billion-parameter model released on 4 August 2025, and two generations are current today. Qwen-Image-3.0 and Qwen-Image-2.1 are not an old and a new version of the same thing. They are two products aimed at two different people.
The hosted flagship
Qwen-Image-3.0 is the model you get when you generate an image through Qwen's own services. Qwen announced it on X in July 2026 as "the third generation of our foundational image generation model". Its headline features are prompts of up to 4.5k tokens, text "legible down to 10px" and native rendering in 12 languages. The launch post sent readers straight to Qwen Chat to try it.
Qwen pitches Qwen-Image-3.0 at dense, text-heavy layouts such as menus, infographics, posters and mock newspaper pages. As of 4 October 2026, Qwen's Hugging Face account carries no Qwen-Image-3.0 weights, so the model only exists as a hosted service.
🎨 Meet Qwen-Image-3.0 — the third generation of our foundational image generation model.
— Qwen (@Alibaba_Qwen) July 22, 2026
If 1.0 was about "Precision," and 2.0 added "Variety, Completeness, Beauty & Authenticity," then 3.0 comes down to a single word: Real (实).
Three dimensions of "Real":
📰 Rich Content — prompts up to 4.5k tokens. One-pass generation of complex layouts: newspapers, storyboards, exam papers — even a 3×3 infographic grid or picture-in-picture-in-picture UIs.
🔬 Authentic Details — text legible down to 10px, full LaTeX paper pages, pores, hair strands & near-photographic skin texture.
🌏 Deep Knowledge — native rendering in 12 languages, 100+ art styles, realistic UIs (web / games / livestreams), plus world knowledge & live web retrieval.
Not just "good-looking" — genuinely useful. Image generation as a real productivity tool for design, content, education & e-commerce.
Go create 🏃🎨
💬Qwen Chat: https://chat.qwen.ai/?inputFeature=t2i
📝Blog: https://qwen.ai/blog?id=qwen-image-3.0
The open model you can download
Qwen-Image-2.1 is the model you can download. Qwen released it on 20 September 2026 on Hugging Face, ModelScope and GitHub at the same time. Qwen-Image-2.1 is one model for both generating and editing, and its visual generator has 7 billion parameters across 32 transformer layers, down from roughly 20 billion in the original Qwen-Image. That drop in size is why people are trying to run Qwen-Image-2.1 on their own machines.
Qwen's announcement on X called it "the most balanced and cost-effective image generation model in the Qwen-Image series":
Meet Qwen-Image-2.1, the most balanced and cost-effective image generation model in the Qwen-Image series! Now open weights! 🎨
— Qwen (@Alibaba_Qwen) September 20, 2026
A unified model for both generation and editing, delivering top-tier quality in a lightweight package.
Highlights: 👀
- Compact & exceptionally fast: A lightweight 7B architecture that outperforms most closed-source models, with drastically accelerated inference for multi-image inputs.
- Native transparency: Natively generates and edits RGBA layers, enabling seamless compositing and text editing within transparent images.
- Versatile, high-fidelity editing: Supports up to 10 reference images and precise local control while preserving strict fidelity for portraits and products.
- Broad coverage & stunning aesthetics: Excels at panoramas, infographics, and virtual try-ons, delivering realistic textures and elegant typography.
Start to create your next masterpiece with Qwen-Image-2.1! 🖼️
- Blog: https://qwen.ai/blog?id=qwen-image-2.1
- GitHub: https://github.com/QwenLM/Qwen-Image-2.1
- Model Scope: https://www.modelscope.cn/models/Qwen/Qwen-Image-2.1
- Hugging Face: https://huggingface.co/Qwen/Qwen-Image-2.1
The table below lines up every Qwen image model a reader is likely to meet, with where each one lives and what it costs.
| Model | Released | Where you use it | Price | Terms |
|---|---|---|---|---|
| Qwen-Image-3.0-Pro | July 2026 | Qwen Chat, Qwen Cloud API | $0.04 per 1K image, $0.075 per 2K image | Hosted only, no weights |
| Qwen-Image-3.0 | July 2026 | Qwen Chat, Qwen Cloud API | $0.03 per image | Hosted only, no weights |
| Qwen-Image-2.1 | 20 September 2026 | Your own computer, Hugging Face demo | Free download | Qwen Research License, non-commercial |
| qwen-image-2.1-pro | 2026 | Alibaba Cloud API | $0.04 per image | Hosted API |
| Qwen-Image (original) | 4 August 2025 | Your own computer | Free download | Apache 2.0 |
API prices are Alibaba Cloud Model Studio's international (Singapore) rates, read on 4 October 2026.
Is the Qwen Image Generator Free?
Yes, there are free ways to use the Qwen image generator, but each one comes with a limit. Which limit applies depends on where you use it.
The Qwen Studio app. Qwen's consumer app, Qwen Studio, is listed as Free on the UK and Singapore App Stores, and its listing names Image Generation as a core feature. The App Store lists the developer as NTH Power Global Tech Singapore, not under the Alibaba name. The web version lives at chat.qwen.ai. Alibaba does not publish a daily image cap for the free app. When we checked on 4 October 2026, an App Store lookup returned no US listing for Qwen Studio, so iPhone owners in the US may not find it.
The API. Developers pay per picture. On Alibaba Cloud Model Studio, Qwen-Image-3.0 costs $0.03 per image at 1K or 2K, and Qwen-Image-3.0-Pro costs $0.04 at 1K or $0.075 at 2K. Edits that send a reference picture add $0.003 per input image. New accounts in the Singapore region get 10 free images, valid for 90 days.
The open weights. Qwen-Image-2.1 costs nothing to download, and Qwen also hosts a free browser demo on Hugging Face. The licence restricts what you may do with the model, which the commercial-use section of this guide covers.
For context, Qwen's chat models follow the same pattern, with a free app and a paid API, which our full breakdown of Qwen's pricing walks through model by model.
How to Use the Qwen Image Generator
The fastest way to use the Qwen image generator is Qwen Chat in a browser: open chat.qwen.ai, pick the image tool and describe the picture you want. Qwen's own launch post for Qwen-Image-3.0 linked to exactly that page. The steps below cover the three official routes.
In Qwen Chat on the web or iPhone
Qwen Chat is the consumer route to Qwen's image models, in a browser or in the Qwen Studio app.
- Go to chat.qwen.ai, or open the Qwen Studio app on iPhone or iPad.
- Choose image generation. Qwen's own launch link, chat.qwen.ai/?inputFeature=t2i, opens Qwen Chat with the text-to-image input already selected.
- Write a detailed prompt. Qwen-Image-3.0 accepts very long prompts, so describe the layout, every line of text you want rendered, the colours and the style.
- Download the result and check any text in it letter by letter before you use it.
If you want Qwen's chat models in a native Mac window rather than a browser tab, our guide to running Qwen on your Mac covers the options.
The free Hugging Face demo
Qwen runs an official Qwen-Image-2.1 demo on Hugging Face, linked from Qwen's own model card. The demo is the quickest way to try the transparent-PNG and multi-image editing features without installing anything. It runs on Hugging Face's shared ZeroGPU hardware rather than on your computer.
Through the API
Developers call Qwen-Image-3.0 through Qwen Cloud or Alibaba Cloud Model Studio. Alibaba's API reference lists an optional Qwen-Image watermark for the corner of each picture, switched off by default. Alibaba's API reference also says generated image links expire after 24 hours, so save every result as soon as it arrives.
Official sites versus lookalike sites
Many of the top Google results for the search qwen image generator are not run by Alibaba. Two of them say so in their own footers. Qwen-image.org states it is "not affiliated with, endorsed by, or sponsored by Qwen, Alibaba Cloud, or any related platform" and sells its own credit subscriptions. Qwenimage21.org calls itself "an independent third-party service". Neither domain appears in Qwen's README or launch posts. These are the official Qwen addresses:
| What you want | Official address |
|---|---|
| Make images as a consumer | chat.qwen.ai and the Qwen Studio app |
| Announcements and blog posts | qwen.ai |
| The paid API | qwencloud.com and Alibaba Cloud Model Studio |
| The open weights | huggingface.co/Qwen, modelscope.cn/models/Qwen and github.com/QwenLM |
A third-party site can still be a legitimate reseller. Just know that you are paying that company, under its own terms, and not Alibaba.
What the Open Qwen Image Model Does Well
Qwen-Image-2.1's standout feature is native transparency: it can output a PNG with a real alpha channel straight from a prompt, with no background-removal step. Transparency is built into the model itself, through an image decoder that handles the alpha channel, so stickers, logos and product cut-outs come out ready to place on any background.
Transparent PNGs and layer editing
Qwen-Image-2.1 can generate a transparent image, edit an existing transparent layer, and pull a subject out of an ordinary photo as a transparent layer. Qwen's GitHub README recommends a fixed prompt wording for transparent output. Copy the line below and replace the middle sentence with your own subject:
This is an RGBA image with transparency. A cartoon fox sticker holding a laptop, thick white outline, flat colours. The image has alpha channel and the background is transparent.
Editing with up to 10 reference images
Qwen-Image-2.1 accepts up to 10 reference images in one request. Qwen's own examples combine six portraits into one group photo and assemble an outfit from five product shots. Another furnishes a whole room from ten furniture photos. You can also mark the area to change with circles, painted marks or a separate mask, and ask for different edits in different regions at once.
A multi-reference prompt works best when it names each image's job:
Image 1 is the person. Image 2 is the jacket. Image 3 is the bag. Show the person from image 1 wearing the jacket from image 2 and carrying the bag from image 3, standing on a rainy city street at dusk. Keep the face, the jacket pattern and the bag logo exactly as they are.
Text inside images
Text rendering is the skill the whole Qwen-Image line is known for. Qwen-Image-2.1 defaults to 2048 by 2048 pixels, so short headlines usually come out crisp. One hands-on test on a MacBook Pro by KGP Talkie found short titles spelled exactly. The long word "LOCALLY" came out as "LOCAILLY", and comic speech bubbles came out as gibberish. Keep text short, and proofread every letter.
How Good Is It? Qwen Image vs Nano Banana and GPT Image
On Qwen's own Qwen-Image-Bench test, Qwen-Image-2.1 scores 60.28 and ranks seventh. That puts it ahead of Google's Nano Banana 2.0 and OpenAI's GPT Image 1.5, but behind GPT Image 2.5 and Qwen's own paid Qwen Image 3 Pro. Qwen publishes the scores only as a chart, so the figures below were read from that chart in Alibaba Cloud's 21 September launch blog.
Qwen-Image-Bench scores
| Model | Qwen-Image-Bench score | Parameters (per Qwen) |
|---|---|---|
| GPT Image 2.5 Sunburst | 67.01 | Undisclosed |
| GPT Image 2 | 64.69 | Undisclosed |
| Grok Imagine 2.0 | 63.47 | Undisclosed |
| Qwen Image 3 Pro | 62.36 | Undisclosed |
| Muse Image | 62.34 | Undisclosed |
| MAI Image 2.5 Pro | 61.02 | Undisclosed |
| Qwen-Image-2.1 | 60.28 | 7B |
| Nano Banana 2.0 | 59.82 | Undisclosed |
| GPT Image 1.5 | 59.65 | Undisclosed |
| Nano Banana Pro | 59.45 | Undisclosed |
| FLUX 2 Max | 55.33 | 32B |
Read the Qwen-Image-Bench table with two caveats. First, this is Qwen's own benchmark, and a vendor's test is not an independent result. Second, the margins are thin: Qwen-Image-2.1 leads Nano Banana 2.0 by 0.46 points. The meaningful finding is that a 7B model you can download lands in the same band as Google's hosted models. Qwen-Image-2.1 also outscores every model on the chart with a published parameter count, including FLUX 2 Max at 32B and Hunyuan Image 3.0 at 80B.
Qwen posted the same chart in its launch thread:
Performance:
— Qwen (@Alibaba_Qwen) September 20, 2026
Public leaderboards
On public leaderboards, Qwen-Image-3.0-Pro entered the Arena Text-to-Image leaderboard at #5 with 1,263 points on 4 August 2026, according to Arena's own post at launch. The previous version, Qwen-Image-2.0-Pro, sat at 1,191 points on the same board. Arena rankings move week to week, so treat that as a launch-day snapshot, not a current position.
Qwen Image vs the generators in your apps
For everyday use, the comparison that matters is with the generators already inside the apps people use. Nano Banana 2 is built into Gemini, and ChatGPT's image generator comes from OpenAI's GPT Image family, which holds the top two places on Qwen's own chart. Qwen's advantage is control, not raw quality: built-in transparency and a model you can download and run yourself.
Can You Use Qwen Image Commercially?
No, not with the free Qwen-Image-2.1 download. Qwen-Image-2.1 ships under the Qwen Research License Agreement, dated 20 September 2026. The agreement defines non-commercial use as "for research or evaluation purposes only" and grants rights for non-commercial purposes only. Commercial use needs a separate licence, which the agreement says you request by emailing Alibaba.
A switch from Apache 2.0
The Qwen Research License marks a reversal. The original Qwen-Image and the December 2025 Qwen-Image-2512 were both released under Apache 2.0, which allows commercial use. Several early reactions flagged the switch, including this post from X user David Hendrickson (@TeksEdge) on launch day:
Qwen’s Image 2.1 released but it come with lots of improvements but just as many caveats.
— David Hendrickson (@TeksEdge) September 20, 2026
The old Qwen-Image-2512 was a 20B model with a massive:
💾 40.9GB BF16 image transformer
But the new Qwen-Image-2.1 comes with …
🧠 7B visual generator
💾 14.2GB BF16
🔥 7.26GB INT8 already available for ComfyUI 👈👀 (Day-1)
This little thing can …
🎨 generate AND edit images
🖼️ generate native 2K
🫥 create real transparent RGBA images
👥 use up to 10 reference images
✏️ preserve people/products while editing
🔤 render text
🎯 do masked/local edits
Qwen basically took the huge local image model and reduced with a catch …
⚠️ The tiny 7.26GB image model comes with an entire pipeline.
It uses a Qwen3-VL 8B encoder + VAE.
ComfyUI already has a 6.31GB W4A8 encoder, and Qwen supports CPU offloading for smaller GPUs.
Also…
🚫 Qwen-Image-2.1 is NON-COMMERCIAL under the new Qwen Research License.
That’s especially notable because Qwen-Image-2512 was Apache 2.0.
What the licence means in practice
In practice, a Qwen-Image-2.1 install on your own machine is fine for testing, learning and research, but not for client work, product photos you sell, or a paid app. The Qwen Research License applies to the downloaded model. Images made in Qwen Chat or through the paid API fall under those services' own terms, so read them before you use an image commercially. Our explainer on whether you can sell AI art covers the wider copyright question.
Can the Qwen Image Generator Run on a Mac?
Qwen-Image-2.1 can run on a Mac, but only through community-built tools, and only on Macs with plenty of memory. Qwen's GitHub README documents setups for Nvidia GPUs, AMD Radeon GPUs and eight other chip platforms through FlagOS, with no Apple silicon, MLX or Metal entry.
Why the 7B model is a 33GB download
The "7B" figure describes only the image generator. Qwen-Image-2.1 also needs a Qwen3-VL 8B text encoder to read your prompt and reference photos. On Hugging Face, the full-precision files add up to about 33GB: 14.23GB for the generator, 17.53GB for the text encoder and 1.35GB for the image decoder. ComfyUI's official package also offers a 7.26GB 8-bit generator and a 6.31GB compressed encoder, which is what makes a laptop install realistic.
What Mac users have measured
Every Mac speed figure so far comes from community tests, not from Qwen. These are the published runs we could trace:
| Mac | Tool | Time per 1024px image | Source |
|---|---|---|---|
| MacBook Pro M5 Max, 36GB | ComfyUI with 8-bit GGUF, 40 steps | 156 seconds | KGP Talkie hands-on test |
| MacBook Pro M5 Max, 36GB | ComfyUI with 8-bit GGUF, full 2048px, 40 steps | 1,094 seconds | KGP Talkie hands-on test |
| M5 Max (64GB is mflux's comfortable default) | mflux (MLX), full precision, peak about 46GB | About 78 seconds | mflux project README |
| MacBook Pro M5, 32GB | Community Core ML conversion | 221 to 250 seconds for the 40 denoising steps | devin-lai Core ML project |
Two lessons come out of those numbers. First, memory is the real limit: KGP Talkie's 36GB machine made new images without swapping, but a two-photo edit at 8-bit pushed ComfyUI to 31GB and wrote 6.5GB to swap. Second, Qwen-Image-2.1's native 2K size is slow on a laptop, so generate at 1024 to 1344 pixels and upscale separately. ModelFit found no published working run on 16GB or 24GB Macs, so a base MacBook Air is not a realistic machine for Qwen-Image-2.1 today.
Execute Automation's walkthrough shows Qwen-Image-2.1 running locally on an M5 Max, including multi-image editing and transparent output:
The easiest local route: ComfyUI
ComfyUI supported Qwen-Image-2.1 from launch day, with official model files and ready-made workflows for text-to-image and editing. On a Mac, testers including KGP Talkie pair it with the ComfyUI-GGUF add-on and a compressed GGUF copy of the model, which lets you pick a size that fits your memory. The Ai Blueprint's tutorial walks through the full setup:
For the bigger picture on downloadable models, including chat models that run comfortably on smaller Macs, see our roundup of the best open-source AI models.
Make Images on Your Mac Without a 33GB Download
Most people want a finished image, not a weekend of ComfyUI setup. Fello AI generates images from a prompt inside a native Mac app using OpenAI's GPT Image, the model family that tops Qwen's own benchmark chart. There is nothing to download beyond the app, no 33GB of model files, and no research-only licence to work around.
The same Fello AI app also gives you Qwen's chat models alongside GPT, Claude, Gemini, Grok, DeepSeek and Kimi. You can draft a detailed image prompt with one model and generate the picture in the same window. Fello AI runs on Mac, iPhone and iPad, has a free tier, and costs $9.99 a month or $79.99 a year for full access.
The Bottom Line
Qwen-Image-2.1 is the highest-scoring downloadable image model on Qwen's own benchmark, and its built-in transparency and ten-image editing make it a strong tool for design assets. It is also the first Qwen image release you cannot use commercially for free. On a Mac, it needs at least 32GB of memory and a few minutes per picture. Try it in the Hugging Face demo for learning and experiments. Use Qwen Chat or the API when you just want Qwen's best images. For paid work, choose a hosted generator whose terms allow it.
FAQ
Is the Qwen image generator free?
Partly. The Qwen Studio app is listed as free on the App Store and includes image generation, and the Qwen-Image-2.1 model is a free download. The API is paid, at $0.03 to $0.075 per image on Alibaba Cloud, and the free download is licensed for research or evaluation only.
Is qwen-image.org the official Qwen image generator?
No. Alibaba's official Qwen addresses are chat.qwen.ai, qwen.ai, qwencloud.com, Alibaba Cloud Model Studio, and the Qwen accounts on Hugging Face, ModelScope and GitHub. Sites like qwen-image.org and qwenimage21.org are third parties reselling access under their own terms.
Can I use Qwen-Image-2.1 for commercial work?
Not under the free licence. The Qwen Research License limits Qwen-Image-2.1 to research or evaluation, and commercial use needs a separate licence from Alibaba. Earlier Qwen-Image releases used Apache 2.0, which did allow commercial use.
Does the Qwen image generator run on a Mac?
Yes, through community tools such as ComfyUI with GGUF files or the MLX-based mflux, but Qwen offers no official Apple silicon support. Published tests show it working on 32GB Macs and up, at roughly 80 seconds to four minutes per 1024-pixel image. Nobody has published a working run on 16GB or 24GB Macs.
Is Qwen Image better than Nano Banana?
Only narrowly, and only on Qwen's own test. Qwen-Image-Bench scores Qwen-Image-2.1 at 60.28 against 59.82 for Nano Banana 2.0. Nano Banana is easier to use inside Gemini, while Qwen-Image-2.1 adds built-in transparent PNGs and the option to run it yourself.

