Simon Willison · 2026-08-16 · notable
Simon Willison — Qwen3.8-27B is excellent, but it overthinks by default
Qwen3.8-27B ships with reasoning effort set to xhigh. Simon Willison measured one SVG prompt taking 21 minutes and 22,276 reasoning tokens on that default; the same prompt with reasoning off finished in 137 seconds.
A strong local model held back by one setting: Qwen3.8-27B arrives with its reasoning effort turned up far too high.
What is it?
Simon Willison tested Qwen3.8-27B, the Apache 2 licensed 27B vision-capable model from Alibaba's Qwen lab, and found it ships with reasoning effort set to xhigh. He rates the size itself highly, calling 27B "an excellent size for running a model on a reasonably specced laptop". The Q4_K_M build is a 17GB file with a maximum context of 262,144 tokens.
How does it work?
Reasoning effort decides how many hidden thinking tokens a model spends before it answers. On the xhigh default, the pelican-on-a-bicycle SVG prompt burned 22,276 reasoning tokens to produce 3,223 tokens of output and took 21 minutes. The same prompt with reasoning switched off returned 3,715 tokens in 137 seconds. Willison ran the official GGUF build in LM Studio on a 128GB M5 Max MacBook Pro and on an NVIDIA DGX Spark, seeing 15 to 30 tokens per second.
Why does it matter?
Anyone downloading this model on the strength of its scores meets the xhigh default first, and a 21-minute wait for a simple drawing reads as a bad model rather than a bad setting. Willison's advice is to "ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first." Same weights, usable speed. He also reports working vision bounding boxes, generated HTML tools and Python scripts, and a successful agent loop.
Who is it for?
developers running local models
Try it
Load the Qwen/Qwen3.8-27B GGUF in LM Studio and set reasoning effort to low