Qwen 3.8 27B is genuinely impressive - if you can stop it thinking for 21 minutes

Alibaba’s Qwen research lab has released Qwen 3.8 27B, an open-weight, vision-capable model small enough to run on a well-specced laptop. Early hands-on testing suggests it’s a genuinely strong performer. There’s just one problem: left on its default settings, it will spend the best part of half an hour deliberating over questions a child could answer in seconds.

What’s actually been shown

Developer Simon Willison, who has been running the model locally on a MacBook Pro and an Nvidia DGX Spark, reports that Qwen 3.8 27B ships out of the box with its “reasoning effort” cranked to the highest of four settings, labelled “xhigh”. Qwen’s own documentation pitches this as the setting for “complex tasks demanding thorough analysis”. In practice, Willison found it applied that same exhaustive process to everything, including drawing a simple SVG of a circle.

His now-famous test case - generating an SVG of a pelican riding a bicycle - took 21 minutes and burned through more than 22,000 “thinking” tokens before producing roughly 3,200 tokens of actual output. Turn reasoning off entirely, and the same model spat out a comparable result in a little over two minutes. Willison rates the “xhigh” version as the best local-model pelican SVG he’s produced, but is blunt about whether the wait was worthwhile: no.

It’s worth being clear about what’s proven here and what isn’t. The overthinking behaviour is directly observed and timed - that’s solid. Qwen’s own benchmark claims that the new model beats its predecessor and even a larger closed-weight Qwen model are self-reported by Alibaba, not independently verified, and Willison flags this explicitly. Take the benchmark charts with a pinch of salt until outside testers weigh in.

Who this actually affects

This is not a story about a broken product that ordinary people need to worry about. Qwen 3.8 27B is an open-weight model aimed at developers and hobbyists running LLMs locally, via tools like LM Studio, rather than something baked into a consumer app or phone. If you’ve never heard of “reasoning effort” settings or quantised GGUF files, you’re not the audience.

For people who do run local models, the practical snag is more mundane than dramatic: the default context window in LM Studio (8,192 tokens) isn’t big enough for the model’s own default thinking style, so early testers hit errors before even getting a straight answer out of it. Willison had to raise the context limit to the model’s full 262,144 tokens just to get it working properly.

What to do about it

If you’re planning to try Qwen 3.8 27B yourself, the fix is simple: don’t leave it on its factory default. Dropping the reasoning effort to “medium” or “low”, or switching it off outright, gets sensible answers in a fraction of the time, without any obvious loss of quality for everyday requests. Commenters on Hacker News broadly echoed Willison’s enthusiasm for how far local, laptop-runnable models have come this year - citing similarly compact models handling image understanding, audio transcription and tool use - while agreeing that Qwen’s default configuration is simply badly chosen for typical use.

The takeaway is mundane rather than alarming: a capable small model has shipped with an over-eager default setting. That’s a tuning quirk fixable with one dropdown menu, not a flaw in the underlying technology - and definitely not something that touches anyone outside the local-LLM tinkering crowd.