Alibaba’s Qwen lab just shipped Qwen 3.8 27B, a 27-billion-parameter, vision-capable model under an Apache 2 license. The headline isn’t the size or the license though. It’s that the model ships with a default setting so aggressive it spends 21 minutes and 22,000 reasoning tokens drawing a pelican on a bicycle, according to Simon Willison, who ran it on both a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark.
Willison calls the default “hilarious,” and he’s right. But underneath the comedy is a real signal about where the open-weights race is heading.
What Actually Shipped
Qwen 3.8 27B runs locally as a 17GB Q4_K_M file. That’s the whole pitch: a strong model that fits on a decent laptop. Willison says it produced the best pelican SVG he’s ever gotten from a locally-run model, with details most models miss, like legs on both sides of the bike and wings that reach the handlebars.
The catch is the reasoning setting. Qwen 3.8 supports a reasoning_effort control with four levels:
- xhigh (the default): thorough analysis for complex tasks
- medium: balances accuracy and speed
- low: optimizes for speed and cost
- off: no reasoning at all
The model ships on xhigh. On consumer hardware, that’s a bad place to start.
The Overthinking Problem, Quantified
Willison’s numbers tell the story better than any benchmark. That 21-minute pelican burned 22,276 reasoning tokens to produce 3,223 tokens of actual output. Turn reasoning off, and the same prompt finished in 137 seconds. Two minutes versus 21, for a result that was arguably just as good.
Then he asked it to “draw an svg of a circle.” The model’s reasoning trace read like an art student talking itself into a thesis project: concentric guide circles, tick marks, a soft gradient, “restrained ambient motion,” a Bauhaus palette debate. Minutes later it delivered a beautiful animated circle study that was not what he asked for.
He also hit LM Studio’s default 8,192-token context limit almost immediately, because the model ate the entire budget just thinking. Bumping context to the full 262,144 tokens fixed it.
Where It Shines
Strip out the overthinking and this is a genuinely capable model. Willison tested its vision skills by asking for bounding boxes around pelicans in a photo, using a 0-1000 coordinate scale. The boxes it returned were a near-perfect match. Even better, he had the model build a full HTML tool to visualize those boxes, offline, from a single prompt. It over-engineered that too, because he forgot to dial down the effort, but it worked.
Why This Matters Now
The industry has spent the last year selling reasoning as the premium feature. More thinking, better answers, that’s the pitch behind every “reasoning model” launch. Qwen 3.8 is a live demo of the limit. Past a point, extra reasoning doesn’t buy you a better circle. It buys you a 21-minute wait and a heating bill.
That’s the bigger trend worth watching. Reasoning has a cost curve, and defaults matter enormously for how people actually experience a model. A great model with a punishing default will get judged on the default. Most users never touch the config.
Practical Takeaways
If you’re running Qwen 3.8 27B locally, Willison’s advice is blunt: ignore the default.
- Start on low reasoning, or turn it off entirely. Dial up only if the output genuinely needs it.
- Load the model with a large context window from the start, or thinking will eat your token budget on trivial prompts.
- Match effort to the task. Bounding boxes, quick SVGs, and simple code don’t need xhigh.
For builders shipping their own models, the lesson is cheaper still: pick a default your users can live with. Willison’s testing shows a model can lose two minutes or 20 to a setting nobody asked to change.
Qwen 3.8 27B is a strong local model wearing a bad first impression. Independent benchmarks will tell us how it stacks up against the closed-weight competition. For now, the takeaway is simpler. The model is worth running. Just don’t let it think itself into a corner. Full details, reasoning traces, and the famous pelican are at Simon Willison’s original writeup.