Qwen 3.8 27B's Default Setting Burns 21 Minutes Per SVG

Qwen 3.8 27B defaults to "xhigh" reasoning effort, burning 21 minutes on a pelican-on-a-bicycle SVG, per Simon Willison's review of Alibaba's model.

The model is Apache 2.0, vision-capable, and quantizes to 17GB, small enough to run locally. That means the 21-minute default is what most people who download it will hit first, not a benchmark footnote.

Willison called it a model that "defaults to wildly overthinking things." Ask it for a plain circle and it starts reasoning about design aesthetics before it draws anything.

Turning reasoning off cuts the wait. Willison's same SVG task dropped to 137 seconds with reasoning disabled, though the output looked visibly worse.

On LM Studio he saw 15 to 30 tokens per second, with Multi-Token Prediction adding about 72% throughput.

Willison's conclusion isn't that the model is broken. Override the reasoning default, and it works for coding, vision, and tool use, he says.

HuggingFace shows 415,000 downloads and 10,500 likes for the base listing. Separately, unsloth's GGUF build alone has pulled 2.73 million downloads.