Alibaba's New Qwen Picture 3 AI Needs to Be Helpful, Not Simply Fairly – Decrypt

Alibaba’s Qwen staff launched Qwen Picture 3.0 on Tuesday, and the pitch has nothing to do with how lovely the output appears to be like. It is about whether or not the output can really be used at work.

Most AI picture instruments—Reve, Nano Banana, Seedream—are designed to excel at particular areas: creativity, realism, modifying capabilities, and so forth. Qwen Picture 3.0 goes in a distinct route. “Qwen-Picture-3.0 is not only pursuing ‘handsome’—it’s pursuing ‘helpful,’ making picture technology a very deployable productiveness instrument,” the Qwen staff wrote within the official announcement.

The centerpiece is what the Chinese language behemoth Alibaba calls wealthy content material. The mannequin accepts as much as 4,500 tokens, which is 4.5 occasions what the earlier technology might course of. Tokens are the items of textual content an AI reads; image a token as roughly one phrase or a part of a phrase, so 4,500 of them are a number of pages of detailed directions

That is sufficient to explain 9 separate infographic panels in a single immediate and get them again as one full picture.

“The complete picture above was generated by Qwen-Picture-3.0 in a single go, slightly than being stitched collectively from a number of pictures,” Alibaba wrote in its weblog. Every panel within the demo comprises its personal diagrams, formulation, captions, and fine-print textual content—rendered in a single shot, not assembled in put up.

That is the one mannequin able to attaining this with out main errors.

The second half is what the corporate calls genuine particulars. Per Alibaba, the mannequin “helps exact rendering of textual content as small as 10px, vividly reproducing particulars like pores and hair strands with lifelike, micro-level depiction.” Ten pixels is ok print—the sort you’ll see on pharmaceutical disclaimers. The mannequin additionally handles LaTeX—the notation system researchers use to write down complicated mathematical equations—precisely throughout full tutorial paper mockups.

In our common assessments we give fashions just a few sentences and consider how they course of them. Qwen Picture 3.0 was capable of generate the picture beneath, per Alibaba’s official weblog.

We tried this function utilizing the mannequin’s quickest configuration. Qwen Picture 3.0 was capable of reproduce one full article from Decrypt. The execution was genuinely spectacular, however the outcome was not flawless.

The third pillar of Qwen Picture 3.0 is deep data. Per the Qwen staff, the mannequin “helps native rendering of 12 languages, simulates mainstream interfaces comparable to internet pages, video games, and livestreams, and attracts on wealthy world data.” It additionally connects to the web to fetch reside information, that means prompting for a climate forecast visible for a particular metropolis and date returns an correct graphic, not a guess.

For instance, Alibaba shared a photograph of an insect on a leaf. The mannequin was capable of generate related textual content based mostly on its understanding of the picture.

Alibaba is pitching design studios, content material groups, e-commerce operations, and educators who want production-ready visible property in bulk.

It’s value noting, although, that in Alibaba’s personal Qwen-Picture-Bench analysis—a benchmark that scores picture high quality, aesthetics, and real-world constancy throughout 18 fashions—Qwen Picture 2.0 Professional, the earlier flagship, positioned fifth. OpenAI’s GPT Picture 2 led the rating. The brand new mannequin could carry out higher, however the launch provides no measured approach to affirm it, as a result of it arrived with out a benchmark desk, downloadable weights, or technical report.

Qwen Picture 1.0 launched with open weights underneath an Apache 2.0 license and a same-day technical report. This one did not. As a part of Alibaba’s latest AI push, the proof right here is solely the hand-picked instance pictures the corporate selected to publish. API trials are open at chat.qwen.ai. Pricing hasn’t been introduced.