Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things
Source
simonwillison.net
Date
Context
Reasoning settings changed the waiting time dramatically in Simon Willison's local tests of Qwen 3.8 27B. In his account, an SVG drawing request took 21 minutes with the default extra-high setting, using 22,276 reasoning tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → before producing its output. The same prompt took 137 seconds with reasoning disabled.
His other example shows why speed alone is an incomplete criterion: a tool generated without reasoning placed image annotations incorrectly, whereas the reasoning-enabled version worked in his test. These are personal experiments, but they illustrate a concrete tradeoff: reducing deliberation can shorten a task substantially while changing whether the result meets its requirements.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
