What it is
Qwen-Image-2.1 was released by Alibaba's Qwen team on September 20, 2026. Its image generator has 7 billion parameters, paired with Qwen3-VL 8B as the text encoder, and one model handles text-to-image, editing with up to 10 reference images, and native transparent (RGBA) images. It generates at 2048 pixels square by default and runs through Diffusers, ComfyUI, SGLang and other common tools.
Unlike earlier Qwen image models, which used Apache 2.0, the weights ship under the Qwen Research License: free for research and evaluation, with commercial use requiring a separate agreement. Alibaba also sells hosted Qwen-Image-2.1-Pro and -Turbo variants, and API-only Qwen-Image-3.0 models.
Our review
Qwen-Image-2.1 is the strongest image model you can download, and you cannot ship it. Artificial Analysis puts it first among open-weights models on its text-to-image arena at 1042 Elo, just ahead of Fibo 2.0 (1039) and Ideogram 4.0 Quality (1026). The Qwen Research License defines permitted use as research or evaluation only, and commercial use needs a separate licence requested by email.
Its lead is narrow at the top of the open field and wide below the closed one. GPT Image 2.5 Sunburst heads the same board at 1222 and Nano Banana 2.1 sits at 1176. Alibaba's own API-only Qwen-Image-3.0-Pro scores 1095, which shows where the company keeps its best work.
The engineering is genuinely useful. One checkpoint handles generation, editing with up to 10 reference images and native RGBA output, with transparency produced by a 64-channel RGBA autoencoder rather than a background-removal pass. The 7B figure covers only the image generator, though. With the Qwen3-VL 8B text encoder the BF16 repository comes to about 33 GB, and Qwen publishes no VRAM figure, offering only CPU offloading for smaller cards.
The licence has two further conditions. Any model trained or fine-tuned on Qwen materials or outputs must say "Built with Qwen" or "Improved using Qwen" in its documentation, and all rights end if the licensee sues Alibaba over intellectual property in the model or its outputs. Evaluation is allowed, so a design-tool company can test it freely; the route to production runs through a sales conversation with no published price.
Where it earns its place
A computer-vision research group at a university
Benchmark and fine-tune a current open image model without API costs
Weights and code are free for research, and the community has already posted 145 adapters. Plan for a large-memory GPU or offloading for the 33 GB checkpoint, and label any released derivative "Built with Qwen".
A design-tool start-up weighing an in-house image backend
Generate stickers, product cut-outs and UI assets with a real alpha channel
Native RGBA removes a matting step and is the reason to evaluate it, which the licence permits. Shipping it needs a commercial agreement with an unpublished price; the hosted Pro variant at about $35 per 1,000 images is the licensed fallback, though whether it exposes RGBA is not stated.
An in-house creative team at a homeware retailer
Compose lifestyle shots from several product reference photos
Editing takes up to 10 references and, per the card, holds product identity across edits. Self-hosted throughput at 2048 px and 40 steps is not published, and production use still needs the commercial licence.
Before you commit
- Setup
- Install Diffusers from GitHub (QwenImage21Pipeline), transformers 5.17+ and torch 2.4+; defaults are 2048 x 2048 at 40 steps, and transparent output needs a specific prompt prefix.
- Integrations
- Diffusers, ComfyUI, vLLM-Omni, SGLang and LightX2V at launch, plus Draw Things and DiffusionBee; two prompt-rewriter checkpoints expand short prompts.
- Data and privacy
- Qwen Research License: research and evaluation only, commercial use by separate agreement, "Built with Qwen" on derived models, termination on IP suits against Alibaba. Self-hosted, so prompts and images stay local.
- Cost at scale
- Hosted Qwen-Image-2.1-Pro is about $35.33 and Turbo $16 per 1,000 images on Artificial Analysis, so 100,000 Pro images a month is roughly $3,500. Self-hosting needs room for about 33 GB of BF16 weights.
- Limits
- No Hugging Face inference provider hosts it, there is no official distilled checkpoint, and it sits at 1042 Elo against 1222 for the closed leader.
What's good
- Top open-weights model on Artificial Analysis (1042 Elo)
- Native RGBA generation and transparent-layer editing
- Edits with up to 10 references and mask or sketch targeting
- Day-one support in Diffusers, ComfyUI and SGLang
What's not
- Research-only licence; commercial terms by negotiation
- About 33 GB of weights once the 8B text encoder is counted
- 180 Elo behind the closed leader
- No published VRAM or speed figures
Best for
Researchers, and teams prototyping image features they will license properly before shipping.
Alternatives
Qwen-Image-2.1: common questions
- Is Qwen-Image-2.1 worth it?
- First among open image models in blind votes, and off-limits for anything you sell until Alibaba signs a separate licence. We rate it 4.0 out of 5.
- How much does Qwen-Image-2.1 cost?
- Weights free for research and evaluation; commercial use needs a separate licence from Alibaba. Hosted Qwen-Image-2.1-Pro is about $35 and -Turbo $16 per 1,000 images, per Artificial Analysis.
- What are the downsides of Qwen-Image-2.1?
- Research-only licence; commercial terms by negotiation; about 33 GB of weights once the 8B text encoder is counted; 180 Elo behind the closed leader; no published VRAM or speed figures.
- Who is Qwen-Image-2.1 best for?
- Researchers, and teams prototyping image features they will license properly before shipping.
- How much does Qwen-Image-2.1 cost for a team or in production?
- Hosted Qwen-Image-2.1-Pro is about $35.33 and Turbo $16 per 1,000 images on Artificial Analysis, so 100,000 Pro images a month is roughly $3,500. Self-hosting needs room for about 33 GB of BF16 weights.
- Is Qwen-Image-2.1 safe for company data?
- Qwen Research License: research and evaluation only, commercial use by separate agreement, "Built with Qwen" on derived models, termination on IP suits against Alibaba. Self-hosted, so prompts and images stay local.
- What are the best Qwen-Image-2.1 alternatives?
- The closest alternatives we have reviewed are Ideogram 4.0, Seedream 5.0 Pro and Nano Banana 2.1.
Answers checked Oct 22, 2026.
Make Qwen-Image-2.1? Show this rating on your site
