Alibaba open source Qwen-Image: 20B parameters, making "text rendering" the home of image models

Most image generation models work hard on "drawing good-looking", but often fail on "writing good characters" - Chinese characters on posters are always short of strokes and few strokes. Qwen-Image, released by Alibaba Tongyi Qianwen on August 4, 2025, chose to tackle the most painful point of this Chinese scene head-on: a MMDiT (Multi-modal Diffusion Transformer) image basic model with 20B parameters, making complex text rendering its home capability.

Focus on "text-intensive" images

The key features of Qwen-Image are highly focused: multi-line typesetting, paragraph-level semantics, fine-grained details and other complex text rendering capabilities, while supporting multiple alphabetical/ideographic languages such as Chinese and English. This means that strong text scenes where "text is the subject" such as posters, logos, and advertisements finally have an open source, Chinese-friendly underlying model. In addition to text rendering, it also supports precise editing as an image base model.

Users can select "Image Generation" in Qwen Chat to experience it directly, and the model is also released as open source on GitHub, Hugging Face, ModelScope and other platforms - domestic developers can download and use it without having to circumvent the firewall, which is a real convenience for the implementation of projects in the Chinese community.

A sustainable open source battle

Qwen-Image is not an isolated action. Fifteen days later (August 19), the team released Qwen-Image-Edit - based on the same 20B base to extend text rendering capabilities to editing tasks. This "generation + editing" dual engine directly competes with international models such as DALL·E, Stable Diffusion, and Flux, and also makes Qwen Alibaba's most representative open source achievement in the direction of multi-modal generation.

From an industry perspective, Qwen-Image's differentiated strategy is quite smart: it avoids the positive consumption of "photography/artistic style" with overseas models, and instead focuses on "Chinese text rendering", a field where overseas models are generally weak but are frequently needed in the domestic design, e-commerce, and advertising industries. For domestic image tool developers, it provides a Chinese base that is commercially available and fine-tuned - this combination of "open source + Chinese scene adaptation" is exactly what China's open source ecosystem needs most.

Several directions worth tracking in the future:

1. Qwen-Image’s community ecology: the number of vertical models (e-commerce, advertising, UI) fine-tuned based on it.
2. The limit of Chinese long text rendering: Whether the stability of ultra-long paragraphs and dense small characters can withstand the test of production.
3. Open source comparison with Flux/SD3.5: How to weigh the dilemma between text rendering and image quality under the same parameter scale.
4. Alibaba’s multi-modal follow-up actions: Whether Qwen-Image will evolve into a part of the video generation base.