Qwen 3.8 27B impresses on local hardware — but its default reasoning mode wastes minutes

Alibaba’s Qwen lab has released Qwen 3.8 27B, an Apache 2-licensed, vision-capable model small enough (a 17GB quantized build) to run on a well-specced laptop or a single NVIDIA DGX Spark. Self-reported benchmarks claim gains over both its predecessor and Qwen’s stronger closed-weight models from earlier in the year, though independent numbers are still pending. Early hands-on testing suggests the model is genuinely capable: it produced the best locally-generated ‘pelican on a bicycle’ SVG the tester had seen and nailed bounding-box detection on photos, returning accurately scaled JSON coordinates.

The headline problem is the shipped default. Qwen 3.8 sets reasoning effort to ‘xhigh,’ which triggers extreme overthinking on even trivial prompts. A single SVG took 21 minutes and over 22,000 reasoning tokens; asked simply to ‘draw a circle,’ the model talked itself into an elaborate animated geometric study nobody requested. The default also blows past LM Studio’s 8,192-token context limit, forcing users to load the full context window just to let it finish thinking. Turning reasoning off cut the same SVG task from 21 minutes to two.

The practical takeaway is to ignore the default and start at ‘low’ or no reasoning, dialing up only when a task genuinely benefits — the tool-building and coding-agent tests showed reasoning does improve one-shot results in some cases. More broadly, it’s a useful reminder that a strong open-weight model can be undermined by a poorly chosen default, and that reasoning-effort controls are now a first-class knob users need to manage themselves on consumer hardware.