Qwen, the AI lab behind the popular Qwen chatbot, just dropped the third generation of its image generation model. And it's not just about pretty pictures — this one can actually read and follow extremely long, complex instructions.
The model accepts up to 4.5k tokens of input — roughly the length of a short essay — and uses that to generate things like newspapers, exam papers, storyboards, and 9-grid infographics. In one demo, Qwen showed a single image containing nine independent pieces of content, all following a prompt that clocked in at about 3.7k tokens.
Text rendering has also gotten a big upgrade. Qwen says the model can generate legible text as small as 10 pixels, and handle LaTeX formulas, superscripts and subscripts, handwritten annotations, and dense newspaper layouts. And it's not just text — details like pores, hair strands, paper texture, and brushstroke textures are also more realistic.
Qwen-Image-3.0 natively supports 12 languages, over 100 art styles, and multiple types of UI interfaces. The team also showed off use cases like generating weather maps from live data, restoring traditional paintings, and creating professional infographics. Compared to previous generations that focused on producing visually impressive images, this one is clearly aimed at practical applications: PowerPoints, educational materials, design work, and content creation.