Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things

Simon Willison tests a 27-billion-parameter open model on his own laptop and documents everything it can do: drive a coding agent through a real codebase, annotate photographs with bounding boxes, generate SVG images, and write working code. He shows how to run it through LM Studio with a 17GB quantised build, why the default reasoning setting is too aggressive and should be turned down, and what multi-token prediction does for speed. The article is the most concrete demonstration available that a local model can now handle agentic workflows without sending data to a commercial service.

Task

Custom workflows

Time

20 min

Level

intermediate

Cost

Free — Nothing to pay. A free account at most.

Tools

Ollama, LM Studio

Published

2026-08-16

Voice

Simon Willison — Explains in plain English what each new model release can and cannot do, tested first-hand rather than repeated from the announcement. Useful to anyone who has to judge a capability claim before committing to it, whether the decision is about a course, a research method, or a procurement case. He also documents failure modes such as prompt injection and unreliable output, which vendor material leaves out.