Alibaba's Qwen launches Qwen-Image-3.0

Alibaba's Qwen team announced Qwen-Image-3.0, the third foundational model in its Qwen-Image series. The team sums up the release in one word, "Real" (实), and says the goal is to make image generation "a truly deployable productivity tool" rather than one that only produces good-looking pictures.

The team groups its claims under three headings:

• Rich content: prompts can now run to 4.5k tokens, which the team says is enough for complex layouts such as newspapers, storyboards and exam papers. Its showcase is a 3×3 grid of infographics produced in one pass, not stitched together from separate images. The grid needed a 3.7k-token prompt and covers subjects from the Sylow theorems in group theory to a medical diagram of chest pain, with Chinese and English text, formulas and charts. A second example nests interfaces inside one another: a VSCode window containing Qwen Chat, which contains WeChat, which contains a coffee poster.
• Authentic details: text as small as 10px stays legible, and skin pores and hair strands are rendered in fine detail. Examples include a page of an algebraic geometry paper with LaTeX-style formulas, a newspaper page, and edits that add handwritten annotations or restore damaged traditional paintings.
• Deep knowledge: the model renders 12 languages natively, can mimic common interfaces such as web pages, games and livestreams, and draws on world knowledge.

The excerpt does not include benchmark results, model size, architecture details, licensing or weight availability. The only access point it mentions is Qwen Chat. All of the capability claims are backed by examples the team chose itself, and none have been independently verified.