Alibaba Launches Qwen-Image-3.0 with 4,500-Token Ultra-Long Prompt Support

Alibaba has released “Qwen-Image-3.0,” the third generation of its Qwen-Image model family. Built around the idea that generated images should be “practical enough to use as a working tool, not just nice to look at,” it emphasizes ultra-long prompt support and high-fidelity text rendering.

Details

• Ultra-long prompt support: Accepts instructions up to 4,500 tokens — roughly 4.5x the 1,000-token cap of Qwen-Image-2.0 — enabling single-pass generation of information-dense layouts like newspaper pages, short-drama storyboards, and dense infographics
• High-fidelity text rendering: Keeps text legible at sizes as small as 10px and reproduces photographic skin and hair texture, with native rendering support across 12 languages and more than 20 fonts
• World-knowledge features: Can reproduce interfaces such as web pages and livestream screens, and claims the ability to pull in current information from the internet during generation
• Target use cases: Newspaper layouts, short-drama storyboards, UI mockups, and e-commerce imagery — positioned squarely at production content work
• A step back on openness: Unlike Qwen-Image 1.0, which shipped under an Apache 2.0 license with open weights and a same-day technical report, this release includes no public benchmarks, no model weights, and no technical report

How to try it

• Available now at chat.qwen.ai; API pricing has not yet been disclosed