Qwen-Image-3.0

Qwen-Image-3.0, released on 21 July 2026, is the point at which Alibaba's image line stops competing on picture quality and starts competing on typesetting. The prompt limit rose from 1,000 tokens in Qwen-Image-2.0 to roughly 4,500, which is enough to describe not a single picture but an entire page - a newspaper spread, a restaurant menu, a storyboard, an exam sheet - and the model is trained to lay that page out in one pass rather than assembling it from separate generations. Text is the headline claim: the team demonstrates legible glyphs down to about ten pixels, mathematical formulae set correctly, twelve languages and a range of typefaces, which is precisely the failure mode that has kept earlier image generators out of design work. Reference-based editing lets an existing picture steer the result, and images can be nested inside images, so mock-ups of screens and printed matter no longer have to be faked in an editor afterwards. Two things are worth stating plainly, because the launch did not. First, this generation shipped with no technical report, no benchmark table and no disclosed parameter count - a reversal for a team whose two previous image models both came with papers and numbers. Second, and more consequentially for anyone who chose Qwen for its openness, the weights were not published: Qwen-Image 1.0 arrived under Apache 2.0 and could be run on your own hardware, while 3.0 exists only as a paid endpoint, listed in Alibaba Cloud Model Studio under the model identifier qwen-image-3.0-pro. Several trackers report the model as an August release for that reason; the API identifier appeared later than the model itself.

▸Videos

No videos yet.

▸Reviews