What this is

A test circulating on Reddit this week caught our eye: Tongyi Qwen 27B (community AWQ-INT4 quantized version) produced a nested animal SVG (scalable vector graphics, common in web illustrations) on two RTX 3090s, consuming 46k tokens per run—with the "thinking" phase alone taking 12 minutes.

The input was a single English prompt: draw a cat riding a zebra, and a zebra riding an elephant, with the output being an HTML file. The author was originally just probing SVG capability (the industry commonly uses "a pelican riding a bicycle" as a benchmark)—and the result exceeded expectations.

What interests us isn't "how well it draws," but the combination: 27B parameters, quantization compression, consumer-grade GPUs, and stable output of structured creative results.

Industry view

Optimists view this as a landmark moment for mid-size open-source models—when creative generation, once considered a moat for large models, can now run locally, the lead of closed-source vendors is being compressed, and the community ecosystem has a new anchor.

We remain skeptical. A single Reddit user test cannot be equated with a benchmark: 46k tokens for one SVG means token cost and electricity aren't trivial; "12 minutes of thinking" is a disaster for any production environment. This looks more like a "single-point surprise" than a "capability leap." Don't rush to treat it as evidence that open-source has caught up with closed-source.

Another overlooked threshold: two RTX 3090s push a full build close to $10K. The real cost here isn't the model—it's the GPU. The so-called "localization freedom" isn't actually cheap for the vast majority of individual users.

Impact on regular people

  • For enterprise IT: Mid-parameter open-source models can serve as local pilot solutions for "specialized generation" tasks—not everything needs to go through cloud APIs.
  • For individual professionals: Designers, front-end developers, and content creators can trial local models on ~$10K hardware for prototyping—validate first, then decide whether to go cloud.
  • For the consumer market: Consumer-grade GPUs are becoming the watershed for "can you run practical AI," and hardware procurement is starting to enter AI tool selection checklists.