Qwen 3.8 27B can do the work. Its default takes forever.

Qwen 3.8 27B fits in a roughly 17 GB quantized file and can annotate images, call tools and write code on capable personal hardware. In Simon Willison’s tests on a MacBook Pro and DGX Spark, the capabilities were impressive. The default reasoning effort was the problem.

A pelican-on-a-bicycle SVG took 21 minutes, with 22,276 reasoning tokens preceding 3,223 output tokens. Turning reasoning off brought another attempt down to 137 seconds, with a worse drawing. A request for a circle became an elaborate animated composition.

The default 8,192-token context could also fill up before an answer arrived. Increasing it fixed that limit, but not the latency. Willison recommends starting with low or no reasoning and increasing it when the task warrants it.

There are trade-offs: without reasoning, his bounding-box tool misplaced the boxes. Serving optimizations helped too; a separate DGX Spark comparison with multi-token prediction was about 72% faster than the LM Studio default. These were individual experiments, not a general performance guarantee.