Qwen 3.8 27B overthinking default
Archived — this story has rotated out of today’s deck. It is kept here in full.
The gist
Qwen 3.8 27B defaults to xhigh reasoning, burning 22,276 tokens on one SVG. Users should run it on low or no reasoning to avoid overthinking.
Background
Alibaba's Qwen lab released Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM, on Friday. It's designed to run on consumer hardware, but its default reasoning effort is set to 'xhigh', causing it to overthink even simple tasks. Simon Willison's review highlighted this issue, showing that a simple SVG request took 21 minutes and 22,276 reasoning tokens, while turning reasoning off produced similar results in 137 seconds.
How it unfolded
- Aug 14, 2026Qwen 3.8 27B is released on Hugging Face, with benchmarks showing improvements over Qwen 3.6 27B and Qwen 3.7-Plus.
- Aug 16, 2026Simon Willison publishes a review noting the model's default 'xhigh' reasoning effort leads to overthinking, using 22,276 reasoning tokens for a pelican SVG that took 21 minutes.
- Aug 17, 2026Discussion spreads on Hacker News and social media, with users echoing the overthinking issue and recommending lower reasoning settings.
Who’s saying what
- Expert
- Simon Willison calls the default 'a hilarious default' and advises running the model on low or no reasoning first.
- Community
- Some users on Hacker News note that overthinking is a speed and cost issue, preferring terse models for agent tasks.
- Skeptic
- A Reddit user claims the overthinking is due to a structural defect in the model's temporal processing layers, not just a default setting.
Still unverified
Claims of a structural defect in SSM layers causing overthinking are from a single Reddit user and not confirmed.