Qwen Image 2.1: Alibaba's 7B Image Generation and Editing Model, Transparency, Turbo, Licence and Pricing

What Qwen Image 2.1 is

Qwen Image 2.1 is an image model from Alibaba's Qwen team, released on 20 September 2026 with open weights on Hugging Face and ModelScope as Qwen/Qwen-Image-2.1. It is a single model for both text-to-image generation and image editing. The Qwen team describes its visual generation component as having 7B parameters in 32 single-stream diffusion transformer layers, and credits mixed-granularity attention and reuse of a prefix key-value cache for keeping compute low. It runs through a dedicated Diffusers pipeline, and vLLM-Omni supported it from release day. On LMArena's text-to-image leaderboard Qwen Image 2.1 has 1227 Elo (rank 21).

Transparency and editing

Qwen Image 2.1 can generate ordinary images or transparent RGBA images directly from text, edit transparent layers, and extract a subject from a photograph, all in the same model; the model card gives a recommended prompt wording for transparent output. Editing accepts up to 10 reference images, and a local edit can be marked with a circle, a painted annotation or a separate mask. The team states that it preserves the identity of people and products across edits, and shows a group photograph built from six portrait references. It also lists better typography, portrait lighting and fine detail as improvements.

Resolutions and running it

The model card lists seven supported aspect ratios at about four megapixels: 2048 by 2048 for square images, 2400 by 1792 for 4:3, 2528 by 1696 for 3:2 and 2752 by 1536 for 16:9, with the portrait versions of each. Its examples use 40 inference steps in bfloat16 on a CUDA GPU, and CPU offloading is offered to lower memory use. Setup needs PyTorch 2.4 or later, Transformers 5.17 or later and the development version of Diffusers. A separate Qwen-Image-2.1-Turbo checkpoint, released on 9 October 2026, uses the same 7B architecture but generates and edits in 8 denoising steps.

Licence

The weights are published under the Qwen Research License Agreement, not an open-source licence. It permits use, copying, modification, derivative works and distribution for non-commercial purposes only, which the agreement defines as research or evaluation. Any commercial use requires a separate commercial licence, which the agreement says can be requested from Qwen by email. This differs from several other Qwen releases, such as Qwen3.8 27B, whose weights are under Apache 2.0. Businesses that want to use Qwen Image 2.1 in a product without a separate licence can instead use the hosted API.

API pricing

The Qwen-Image-2.1 Pro and Turbo APIs went live on Alibaba Cloud Model Studio on 9 October 2026, and both are billed per output image. On the international deployment in Singapore, qwen-image-2.1-pro costs $0.04 per image and qwen-image-2.1-turbo costs $0.016 per image, and each comes with a free quota of 10 images. Global deployments in China (Beijing), Hong Kong, Frankfurt, Virginia and Tokyo list $0.035333 per image for Pro and $0.014133 for Turbo. The earlier qwen-image-2.0-pro is listed at $0.075 per image internationally.

Official Resources