Alibaba has launched the Qwen-Image-3.0 model series. In blind evaluations conducted by Artificial Analysis, the Pro version jumped from 34th to 6th in image editing ranking, and rose from 12th to 9th for text-to-image generation. An 83 Elo gain for image editing and 48 Elo gain for text-to-image are massive improvements for any model.

Qwen-Image-3.0 comes in two variants: the Pro version is the flagship release featuring agent-based prompt rewriting, which automatically optimizes user prompts; the standard edition delivers comparable capability with faster inference speed, making it ideal for batch generation. The models prioritize photorealism and information density: they support 4.5k-token prompts, can render text as small as 10 pixels, and natively support 12 languages. On Alibaba Cloud Model Studio, the Pro version is priced at $0.04 per 1K resolution image and $0.075 per 2K resolution image; the standard edition costs $0.03 per image for both resolutions. For comparison, GPT Image 2 high, the top-ranked model on the leaderboard, costs $211 per 1,000 images. The standard edition ranks 11th for text-to-image and 15th for image editing, with Elo gains of over 135 and nearly 100 respectively compared to the previous generation.

On Artificial Analysis' Text to Image Leaderboard, Qwen-Image-3.0-Pro ranks 9th with 1282 Elo, behind Nano Banana 2 Lite and ahead of Seedream 5.0 Pro. On the Image Editing Leaderboard, it ranks 6th with 1249 Elo, following models including MAI-Image-2.5-Pro, Reve 2.1, and GPT Image 2. Check the full rankings in the video below:

Below are three text-to-image examples. The first is an infographic design requiring precise layout and multiple clock labels; the second is an aerial garden floor plan demanding high detail; the third is a four-panel comic requiring coherent narrative and consistent character expressions. The results confirm that Qwen-Image-3.0-Pro delivers strong performance in text rendering and layout control.




For image editing tasks, it can accurately add text and elements to existing images. Examples include adding labels to a fish, inserting containers into a freight scene, and adjusting the camera angle of a game scene. All these operations require understanding of spatial relationships and consistent preservation of the original image's style.




One netizen commented: "An 83 Elo increase is a huge leap, but Qwen keeps rolling out new image models and no one even notices." It does feel rather under the radar. While all the attention is on video generation right now, image models are iterating at an astonishing pace. What's more notable is that most of this performance gain comes from engineering optimization and prompt rewriting. Additionally, priced at just a fraction of top-tier competing models, it makes high-quality image generation accessible to far more users. Another netizen mentioned they would choose this model over older alternatives for any project requiring static images.

If you want to test its performance yourself, you can vote on it at Artificial Analysis' Image Arena, or try it directly for free on Qwen Studio.

[Text to Image Leaderboard](https://artificialanalysis.ai/image/leaderboard/text-to-image)

[Image Editing Leaderboard](https://artificialanalysis.ai/image/leaderboard/editing)

[Image Arena](https://artificialanalysis.ai/image/arena)