Qwen-Image-3.0: photographic text rendering, no weights or API
Alibaba's new image model handles 4.5k-token prompts and nested UIs in Qwen Chat today — yet ships with no downloadable model and no benchmarks.
Copy markdown
What it does, and where to try it
Qwen-Image-3.0 takes prompts up to ~4,500 tokens and renders text down to roughly 10px legibly — LaTeX with correct sub/superscripts, dense newspaper columns, and 12 languages in their real scripts. It also holds nested UI screenshots (a VS Code window inside a chat inside a poster). It's live now in Qwen Chat on web and mobile, with usage limits on free accounts.
The catch: no weights, no API price, no benchmarks
Unlike Qwen-Image 1.0 and 2.0, the 3.0 release ships with no downloadable weights, no model card or technical report, no benchmark table, and no published API pricing — it's cloud-only in Qwen Chat. If you self-host Qwen's image models or planned to build on the API, there's nothing to pull yet.
Alibaba's split: open on text, closed on images
It lands days after Alibaba previewed Qwen3.8-Max (2.4T params) as API-first, with open weights promised "soon." The pattern for builders: Qwen's language line is staying open while its image line goes hosted-only — worth planning around if Qwen-Image weights are in your stack.