Qwen's New Image Model Ships Closed, With NSFW Tags Buried in Its HTML

Qwen's pitch: full pages, not just pretty pictures

Alibaba's Qwen team put out Qwen-Image-3.0 on its own blog under the tagline "Rich Content, Authentic Details, Deep Knowledge." The demo images leaned hard on that promise: full newspaper spreads, multi-panel manga chapters, storyboards, exam sheets with worked equations. The post's own spec line, quoted in the thread, put a number on it: a single prompt can run "up to 4.5k token input, effortlessly generating complex layouts such as newspapers, storyboards, and exam papers." The blog singled out three language demos, Japanese, Korean and Spanish, as proof it finally renders non-Latin text without garbling it.

That's the entire pitch, and it's a real gap in the field: not a smarter single image, but a model that can hold a whole page, grid, captions, running text, coherently. It's exactly the kind of layout task that breaks older diffusion models, which is why the claim traveled so fast.

Qwen's LLMs, and its earlier image models, built their name by landing on Hugging Face with open weights on release day. Not this time. Multiple commenters combed the post for a release date or license and came up empty: no weights, no timeline, nothing. It's the kind of absence that only reads as news because of what usually happens instead.

The page's own metadata says more than the blog does

A commenter with 3,600+ karma didn't stop at the demo images. They checked the page's source instead.

"The meta keywords in the HTML is very interesting. 100+ references to NSFW topics such as hentai, nudes, etc." — a longtime HN user who read the launch page's raw HTML

Nobody in the thread explained why a product blog would carry that many adult-content keywords baked into its metadata: leftover SEO cruft, an artifact of training-data tags, or something else entirely. The finding sat there, unresolved and unaddressed by Qwen.

A color cast that pointed at a rival's model

"They must have trained on GPT Image 1 outputs. The yellow tint is unmistakable." — a developer building LLM systems

The commenter posted several of Qwen's own demo outputs alongside known GPT Image 1 (OpenAI's image generator) samples, matching a color signature that shows up across unrelated prompts: the kind of tell that's hard to fake and easy to spot once you've seen it once.

When the demo credits a dead artist

The manga-style demo carried Japanese production credits: "原作・監修 三浦建太郎" (original work, supervised by Kentaro Miura) and "共同制作 白泉社" (co-produced with Hakusensha), a real publisher, and an artist who died in 2021.

"It's a bit shocking to see them showing off blatantly disinformation-al examples... I hope these stakeholders have good enough layers." — a commenter who read the Japanese captions in the demo closely

Native text rendering, checked by someone who reads the language

The blog's own language showcase claimed the model "accurately renders Japanese, Korean, and Spanish." A longtime systems programmer, creator of the Chrono Rust library and on HN since 2010, went through the Korean sample character by character and found it wasn't close:

"And yet the Korean text is not accurate... "드레스 컬렉션 dress collection" has vowels ㅔ mixed with ㅐ, "초웜한" should be "초월한 exceeding"... and so on." — lifthrasiir, creator of the Chrono Rust library

Other testers who moved past the blog's curated set reported the same drop-off once they typed their own prompts instead of the demo ones: a Polish GDP chart came back as "slop," and a separate test turned up third legs and glowing eyes, the sort of failures the marketing images had none of. Simon Willison, creator of Datasette, put his finger on why the demos read as more polished than the model itself: the post's headline example needed a 3.7k-token prompt to pull off, and the blog never showed that prompt.

Qwen just swapped sides in the openness argument

For a couple of years the contrast on HN has been simple: OpenAI and Google keep their weights locked, Qwen ships them. Qwen-Image-3.0 quietly drops that contrast, and the clearest evidence for what changed isn't in a press release. It's sitting in a page's own metadata, and in Korean vowels a reader had to fix by hand.