What he said
Simon Willison spent the weekend running Qwen 3.8 27B, a new AI model from Alibaba’s Qwen research lab, on his own laptop and a dedicated AI workstation. It’s “open-weight,” meaning anyone can download the trained model file itself and run it on their own hardware, instead of only accessing it through the vendor’s paid app or API. The “27B” in its name means it has 27 billion parameters, the internal values a model learns during training; fewer parameters generally means a smaller file that needs less computer memory to run, which is why this size is meant for a personal laptop rather than a data center. His verdict on how the model ships out of the box: “This is a hilarious default. It’s absolutely not a good way to run the model, especially on consumer hardware.”
The setting he’s describing is the model’s reasoning effort, how much internal step-by-step thinking it does before answering, which trades speed for (sometimes) accuracy. Qwen’s own documentation lists three levels, low, medium, and xhigh, and says the model defaults to xhigh, “for complex tasks demanding thorough analysis.” Willison’s recommendation is blunt: “ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first.”
Who he is
Willison is an independent open source developer. His about page describes him as working full-time building open source tools for data journalism, centered on Datasette, a tool for exploring and publishing datasets. He also created llm, a widely used open source command-line tool for working with AI models, and co-created the Django web framework, a widely used tool for building websites. He writes and publishes hands-on tests of new AI models within days of their release, which is why builders treat his benchmarks as a fast, credible second opinion.
What he gets right, and where it’s thin
Willison’s evidence is concrete, not a vibe check. Asked to draw a simple SVG (a vector image format websites use for icons and diagrams) of a bicycle with a pelican riding it, a running benchmark he uses to compare models, the model spent 21 minutes and burned through 22,276 reasoning tokens (the model’s internal “thinking” text, which costs money and time even though you never see it) before producing an answer built from 3,223 output tokens, the part you actually see. With reasoning turned off, the same prompt finished in 137 seconds. Asked for something as basic as “draw an SVG of a circle,” the model reasoned at length about adding concentric guide circles, tick marks, gradients, and animations nobody asked for.
That’s a specific, reproducible finding: on identical hardware, the default setting turns a two-minute job into a 21-minute one for no accuracy gain Willison observed. What his post doesn’t do is quantify how much better xhigh performs on harder tasks that actually need deep reasoning, so it’s a strong caution about the default, not a full verdict on whether xhigh is ever worth the wait.
Why it’s notable
Qwen 3.8 27B is sized to run on a well-specced laptop rather than a data center, which is exactly the audience most likely to feel a 21-minute wait as a real cost, not a rounding error on a cloud bill. An open-weight model’s default settings matter more than a hosted one’s, because nobody is optimizing your bill or your wait time for you; you’re running the raw software as the vendor shipped it.
What it means for builders
If you download an open-weight model to run locally or on your own server, check its default reasoning or “effort” setting before you build anything on top of it, the same way you’d check any tool’s factory settings before trusting them. A vendor’s default is often tuned for benchmark scores, not for your wait time or your compute bill. Start at the lowest reasoning level, confirm the output quality is good enough for your task, and only turn up the reasoning effort for the specific requests that actually need it.
End of article