Qwen-Image

A 20-billion-parameter image model

Qwen-Image is an image generation foundation model with 20 billion parameters, released on 4 August 2025 together with a technical report, its weights and a blog post^[1]. A foundation model is not built for one narrow task but can be adapted to many; the parameters are the numbers learned during training that decide what the model draws for a given prompt. Models of this kind typically start from random noise and, step by step, shape it into an image under the guidance of a text description.

The scale shows up directly in the hardware. The model card lists Qwen-Image at 20 billion parameters in BF16^[3], a 16-bit floating-point format. Storing that many parameters in BF16 takes tens of gigabytes before any working memory for producing an image is counted. The weights are hosted on Hugging Face, and Qwen-Image demo spaces are among the Qwen organization's public demos there^[2].

Text rendering, especially Chinese

What made Qwen-Image stand out is how it handles text inside generated images, especially Chinese^[1]. The model card describes it as rendering alphabetic languages such as English and logographic scripts such as Chinese^[3]. Letters and characters are unforgiving: one stroke out of place turns a word into a different word or into nonsense, which is why earlier image models often produced letter-like scribbles. The two kinds of writing stress a model differently. Alphabetic scripts are built from a small set of reusable letters, while Chinese characters are thousands of distinct whole shapes, many with a fixed internal arrangement of strokes.

Output sizes and licence

The Quick Start in the model card uses 1328x1328 for 1:1 output, 1664x928 for 16:9 and 1472x1140 for 4:3^[3]. These are aspect-ratio presets: rather than always generating a square and cropping it, the model is run at resolutions that already match common frame shapes. The weights are released under Apache 2.0^[1], a permissive open-source licence that allows use and modification, including commercial use, as long as the licence terms are kept.

Editing, and later checkpoints

Qwen-Image is not only a text-to-image generator. It also supports image editing, and later editing and generation checkpoints included Qwen-Image-Edit-2511 and Qwen-Image-2512^[1][2]. The naming keeps editing models apart from plain generation models, so a user can pick the checkpoint that matches the task at hand.

Place in the Qwen line

The line kept moving after the first release. Qwen-Image-2.1, a 7-billion-parameter text-to-image model, and a Qwen-Image-2.1-Turbo variant appeared on Hugging Face in September–October 2026^[2]. The newer generation is far smaller than the original 20-billion-parameter model while keeping the Qwen-Image name, which suggests the family has been extended toward lighter, easier-to-run models alongside the large one.