Qwen-Image-2.1
Alibaba's 7B image model that generates, edits and makes real transparent PNGs from up to 10 references. Research-only license.
In short. Qwen-Image-2.1 is an AI image model from Alibaba's Qwen team. You can create pictures from text, edit photos with instructions, combine up to ten reference images, and generate images with a genuinely transparent background, which is handy for logos, stickers and product cutouts. The weights are free to download, but the license only allows research and evaluation, so business use needs a separate agreement.
Summary
Alibaba's Qwen team open-sourced Qwen-Image-2.1 on 20 September 2026, and added an accelerated Qwen-Image-2.1-Turbo checkpoint that needs only 8 denoising steps on 9 October 2026. One model handles text-to-image generation and image editing. Its visual generation component has about 7 billion parameters: a single-stream diffusion transformer with 32 layers that samples with flow matching. The text encoder is Qwen3-VL 8B, a separate 8-billion-parameter vision-language model that reads both the instruction and any reference images.
The headline feature is native transparency. A 64-channel autoencoder produces an alpha channel, so the model can generate transparent (RGBA) images directly, edit transparent layers, and extract a subject from an ordinary photograph as a cut-out. Other capabilities include up to 10 reference images in one request, such as group portraits or virtual try-on, and local edits specified with circles, painted annotations or separate masks. Generation is native 2K (for example 2048 by 2048), with improved text rendering and portrait lighting.
The weights are in Diffusers format (the QwenImage21Pipeline), and day-zero support was announced for Diffusers, ComfyUI, vLLM-Omni, SGLang, LightX2V and OpenVINO. The generator is about 14 GB in BF16, but the full pipeline with the Qwen3-VL text encoder is about 33 GB, so most people use CPU offloading or the INT8 repack for ComfyUI. Qwen also publishes two Qwen3.5-VL 9B prompt-rewriting models, one for generation and one for editing, and recommends using them to turn short prompts into detailed ones.
The license is the Qwen Research License Agreement, not Apache 2.0 as for the earlier Qwen-Image models: research and evaluation are free, and any commercial use needs a separate license from Qwen. A Qwen-Image-2.1 Pro variant, and hosted versions of Turbo, are offered only as APIs through Alibaba Cloud Model Studio. Qwen's claim that the model beats most closed models comes from its own Qwen-Image-Bench and has not been independently verified.
Good for
- Transparent assets such as logos, stickers, icons and product cut-outs, straight from the model
- Editing with several references, for example group portraits, virtual try-on or room layouts
- Targeted edits using circles, annotations or masks instead of describing the area in words
- Research and evaluation of image generation and editing
Watch out for
- The license is research and evaluation only; commercial use needs a separate license from Qwen
- The full BF16 pipeline is about 33 GB because of the 8B text encoder; plan for CPU offloading or the INT8 repack
- Recent software is required (Transformers 5.17 or newer and Diffusers from source, per the model card)
- The Pro variant is API-only and not part of the open weights
- Benchmark results are Alibaba's own and have not been independently checked
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Qwen-Image-2.1. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Qwen-Image-2.1 (40 steps) BF16 | 7.1B | 33 GB | 38 GB | Workstation |
| Qwen-Image-2.1-Turbo (8 steps) BF16 | 7.1B | 32 GB | 37 GB | Workstation |
| INT8 (ComfyUI repack) INT8 | 7.1B | 17 GB | 20 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
Alibaba's research-only license. You may use, modify and share the model for research or evaluation, but any commercial use needs a separate license from Qwen. Alibaba's hosted API is the usual commercial route.
- Attribution or model name must be shown
More about Qwen Research License
See also
- FLUX 3 Image Klein vs Qwen-Image-2.1: licenses, memory and features
- Best open-source image and video generation models
- Cosmos 3: NVIDIA's open world models for robots and driving
- Krea 2: an open image model trained from scratch for variety
- MOSS-Image: a 5B open image model for long text and JSON layouts
- Qwen3.8: three language models, three licenses, and a fourth for images
Frequently asked questions
- What is Qwen-Image-2.1?
- Qwen-Image-2.1 is an AI image model from Alibaba's Qwen team. You can create pictures from text, edit photos with instructions, combine up to ten reference images, and generate images with a genuinely transparent background, which is handy for logos, stickers and product cutouts. The weights are free to download, but the license only allows research and evaluation, so business use needs a separate agreement.
- Can I use Qwen-Image-2.1 commercially?
- Not without separate permission. Qwen-Image-2.1 is released under the Qwen Research License, which allows research and personal use only.
- Is Qwen-Image-2.1 open source?
- Qwen-Image-2.1 is open-weight, not open source in the strict sense. You can download the model files, but the Qwen Research License adds conditions that the Open Source Initiative definition does not allow.
- Can I run Qwen-Image-2.1 on my own computer?
- Yes, on a good graphics card. The smallest variant, INT8 (ComfyUI repack), needs about 20 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Qwen-Image-2.1 released?
- Qwen-Image-2.1 was released by Alibaba Qwen on September 20, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Qwen-Image-2.1 repository and release log (GitHub) github.com
- Qwen-Image-2.1 model card (Hugging Face) huggingface.co
- Qwen-Image-2.1-Turbo model card (Hugging Face) huggingface.co
- Alibaba Cloud blog: Qwen-Image-2.1, compact, efficient and unified image creation alibabacloud.com
- ComfyUI blog: Qwen-Image-2.1 in ComfyUI (20 Sep 2026) blog.comfy.org
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| FLUX 3 Image Klein Black Forest Labs | 5.6B | – | FLUX Kommunity (Conditional) | 8 GB Laptop or phone | |
| LTX-2.5 LTX (Lightricks) | 22B | – | LTX-2.x Community (Conditional) | 16 GB Laptop or phone | |
| Z-Image Tongyi MAI | 6.15B | – | Apache 2.0 (Permissive) | 16 GB Laptop or phone | |
| MOSS-Image OpenMOSS | 5.08B | – | Apache 2.0 (Permissive) | 4.2 GB Laptop or phone |