Qwen Image 2.1 is Alibaba's open model for generating and editing images. The visual part has 7 billion parameters, and it offers native 2K output, editing from ten references and a true alpha channel. The weights are free to download, but the license allows non-commercial use only. Below: versions, ways to run it, a memory estimate and working alternatives.
What Qwen Image is and how the versions differ
Qwen Image is not a single model but a whole lineup from the Qwen team at Alibaba. The timeline, taken from the official repositories:
| Version | Date | What is known |
|---|---|---|
| Qwen-Image | August 4, 2025 | 20B parameters, Apache 2.0 license, weights on Hugging Face, strong at text in English and Chinese |
| Qwen-Image-Edit-2511 and Qwen-Image-2512 | December 23 and 31, 2025 | Updates for editing and generation, weights on Hugging Face and ModelScope |
| Qwen-Image-2.0 | February 10, 2026 | Announcement of the new base model, with a try-out offered in Qwen Chat |
| Qwen-Image-2.1 | September 2026 | 7B, open weights under the Qwen Research License |
| qwen-image-2.1-turbo and qwen-image-2.1-pro | API from October 9, 2026 | Paid versions in Alibaba Cloud Model Studio |
The table draws on the README of the QwenLM/Qwen-Image repository, the Hugging Face model cards and the Alibaba Cloud blog. The card gives no exact release date for 2.1; press reports say September 20, 2026.
The main difference between the first model and 2.1 is not size but license. The original Qwen-Image shipped under Apache 2.0 and was fine for a product. Version 2.1 came out under the Qwen Research License. If you planned to build it into a commercial service, that changes everything.
What Qwen-Image-2.1 can do
What we know from the Hugging Face model card and the Alibaba blog:
- It draws from text straight at native 2K, 2048x2048 by default. The card lists the aspect ratios 1:1, 4:3, 3:4, 3:2, 2:3, 16:9 and 9:16.
- It edits finished images from local selections (circles, marks, masks), takes up to 10 references in one request and keeps the look of people and products.
- It supports a transparent background: it returns an image with an alpha channel directly from a text prompt, edits transparent layers and cuts an object out of a photo.
- The API has a Turbo variant with the same architecture but 8 steps instead of 40.
On text inside images, Alibaba says headlines and captions are legible and dense layouts such as infographics hold together. Those are the company's own words, and the sources we read contain no independent measurements. The first Qwen-Image claimed English and Chinese. The 2.1 materials say nothing about Cyrillic, so test Russian lettering yourself.
Here is a ten-minute test. Take three short phrases in Russian (a shop sign, a poster headline, a packaging label) and run each five times with different seeds. Count the share of runs where the text comes out with no missing or extra letters. If fewer than four out of five are clean, do not rely on this model for Cyrillic.
Where to use Qwen Image 2.1
| Method | What you need | Limits |
|---|---|---|
| Hugging Face and diffusers | A GPU, Python | Non-commercial use only |
| ComfyUI | The template from the ComfyUI docs, model files | The same license terms |
| Alibaba Cloud Model Studio | An Alibaba Cloud account, payment | Turbo $0.016 and 120 requests per minute, Pro $0.04 and 20 requests per minute |
| Qwen Chat | A browser | Official sources do not confirm that it runs 2.1 |
| codexkey | A sk_cdx_... key | No Qwen Image in the catalog |
API prices and limits come from the qwen-image-2.1-pro page and the Alibaba blog. This is the Singapore price list for October 2026 without promotions, and other regions show different numbers. You pay for each successful image. We did not check whether Alibaba Cloud registration and payment work from Russia.
Read the license before you run anything. In the Qwen Research License text, non-commercial use is defined as "research or evaluation purposes only", and any commercial use requires a separate license from Alibaba. There is no revenue or user threshold: the restriction is tied to purpose. If you distribute models built on these weights, you must credit "Built with Qwen" or "Improved using Qwen".
How to run Qwen-Image-2.1 locally
Installation per the model card:
pip install "torch>=2.4.0" "transformers>=5.17" accelerate pillow
pip install git+https://github.com/huggingface/diffusers
A minimal generation, following the card's example:
import torch
from diffusers import QwenImage21Pipeline
pipe = QwenImage21Pipeline.from_pretrained("Qwen/Qwen-Image-2.1", torch_dtype=torch.bfloat16)
pipe.to("cuda")
image = pipe(
prompt="A bookshop sign reading 'Chitalka' above the door, evening light",
width=2048,
height=2048,
num_inference_steps=40,
generator=torch.Generator("cuda").manual_seed(42),
).images[0]
image.save("sign.png")
When you run short of video memory, the card suggests pipe.enable_model_cpu_offload(). Check argument names against the current README, because diffusers changes fast.
Memory for the weights comes from the formula "parameters × bytes per weight". The visual part is 7B. The text encoder, judging by the file names in the ComfyUI template, is Qwen3-VL 8B.
| Precision | Bytes per weight | Visual part 7B | Encoder 8B | Total for weights |
|---|---|---|---|---|
| bf16 | 2 | 14 GB | 16 GB | about 30 GB |
| int8 or fp8 | 1 | 7 GB | 8 GB | about 15 GB |
| 4-bit (GGUF Q4) | 0.5 | 3.5 GB | 4 GB | about 7.5 GB |
This counts weights only, without the VAE and activations, so real usage is higher. For orientation, Unsloth estimates 11 GB of video memory with GGUF and 24 GB with INT8 or FP8, and says itself that these are not verified minimums. On smaller cards part of the weights has to be offloaded to system RAM. Unsloth allows a run with 6 GB using offloading but warns that it will be slower.
Qwen Image on codexkey: what is missing and what to use instead
As of October 29, 2026, the image models in the codexkey catalog are nano-banana, nano-banana-2, nano-banana-pro, gpt-image-2 and gpt-image-2-5. Qwen Image is not among them, and you cannot reach it through our key. If you need a working API from Russia, these are the closest substitutes:
| Task | What to take | Where to read more |
|---|---|---|
| Dense text, infographics, 2K and 4K | nano-banana-pro | Nano Banana Pro review |
| Cheap drafts | nano-banana-2 with "tier": "lite" or nano-banana-pro on the Cheap channel | Nano Banana for free and on a budget |
| General generation and editing through the OpenAI API | gpt-image-2, gpt-image-2-5 | GPT Image 2 and 2.5 |
A request to gpt-image-2 looks like this (an example from the documentation):
curl https://api.codexkey.online/v1/images/generations \
-H "Authorization: Bearer sk_cdx_YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "gpt-image-2", "prompt": "a red apple on a white table", "resolution": "1K", "size": "1024x1024"}'
The public price list has no per-image prices; they are on the model page in the catalog. Parameters and the Premium and Cheap channels are described in the images and video reference. For choosing an image API by task, see the comparison GPT Image and Nano Banana.
When Qwen Image fits and when it does not
Take it for research, study and internal prototypes on your own graphics card: the weights are open and nobody counts your runs. It also helps when you need a transparent background and object cut-outs from a single model.
We would not build a commercial product on it without a separate agreement with Alibaba, because the license explicitly forbids using the weights commercially. Alibaba does have a paid API, Model Studio, but we did not check access to it from Russia. Skip the model if Russian text in the image is critical and you have no time to test: the official materials say nothing about Cyrillic.
Do not pick between Qwen-Image-2.1, nano-banana-2 and gpt-image-2 from someone else's rankings. Put together 15 real tasks: five with text, five with references, five with a complex composition or a transparent background. Run each one three times on each model. Count what share of results go into work without manual rework, and divide your spend by the number of usable images. An attempt costs different amounts on different models, and what decides it is the price of one usable image.
For the Qwen text models in our catalog, read the Qwen 3.8 review and the Claude vs Qwen comparison.
FAQ
What is Qwen Image?
It is a lineup of Alibaba (Qwen team) models for generating and editing images. The first Qwen-Image came out on August 4, 2025 with 20 billion parameters under Apache 2.0. The current version, Qwen-Image-2.1, has 7 billion parameters and open weights under a research license.
Is Qwen Image 2.1 free?
The weights can be downloaded from Hugging Face for free, but you may use them only for research and evaluation. Alibaba's API is paid: Turbo costs $0.016, Pro $0.04 per image, per the Singapore price list for October 2026.
Can I use Qwen Image 2.1 in a commercial project?
Only under a separate commercial license from Alibaba: the Qwen Research License explicitly forbids commercial use without one. For production you are left with the paid Model Studio API or models available through a gateway such as codexkey.
Is Qwen Image available on codexkey?
No. As of October 29, 2026 the catalog has no Qwen Image models. For images you can use Nano Banana, Nano Banana 2, Nano Banana Pro, GPT Image 2 and GPT Image 2.5.
How much video memory does Qwen Image 2.1 need?
The card gives no minimum. By the "parameters × bytes per weight" formula the weights take about 30 GB in bf16, 15 GB in int8 and 7.5 GB in 4-bit quantization. Unsloth estimates 11 GB with GGUF and 24 GB with INT8 or FP8.
Does Qwen Image draw Russian text well?
We found no confirmation: Alibaba's materials on 2.1 do not name languages, and the first model claimed English and Chinese. Test it on your own phrases, five runs each, and count the share of clean results.