Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things
Simon Willison’s Weblog Simon Willison ● Covered by 2 sources
Opinion — commentary, not a factual news event.
Qwen 3.8 27B is out, and its default setting makes it think way too hard. That’s funny until it eats minutes, but the model itself is strong.
Based on reporting by Simon Willison’s Weblog, Simon Willison — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Alibaba’s Qwen lab has released Qwen 3.8 27B, an Apache 2 licensed vision-capable model with 27 billion parameters. Simon Willison went into it expecting a lot: the size is small enough to run on a well-specced laptop, and its predecessor, Qwen 3.6 27B, had already been impressive. Qwen’s own benchmark claims point to a clear jump, including a result that beats both Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus, which had been one of the lab’s strongest models as recently as May.
The catch is the default reasoning setting. Qwen 3.8 ships with xhigh reasoning effort turned on, and Willison found that to be a terrible starting point for normal use. On consumer hardware, the model happily burned through context on trivial prompts. With LM Studio’s default 8,192-token limit, it could chew up the whole window just thinking. After switching to the full 262,144-token context, it produced a pelican-on-a-bicycle SVG, but only after 21 minutes and 22,276 reasoning tokens. Turning reasoning off brought that same job down to a little over two minutes.
The overthinking wasn’t just annoying; it also changed the model’s behavior in obvious ways. Ask it for a simple circle SVG at xhigh and it starts imagining a “circle study” with gradients, motion, and design choices the user never requested. But the model can also be accurate when the task needs it. On a pelican photo, it returned two bounding boxes that matched the birds well. And when Willison used it to build an offline HTML tool for labeling boxes, it even made itself a demo scene he didn’t ask for, apparently because the example JSON mentioned pelicans.
That same tendency showed up in coding. Through Pi, Qwen 3.8 27B was able to inspect auth code, then write and test a Python script to convert a JSONL transcript to markdown. The model also has enough long context, tool calling, vision ability, and code generation to act like a real local coding agent. The problem is speed. Willison says he was seeing about 15 to 30 tokens a second in LM Studio, which is usable but not exactly brisk compared with hosted APIs.
There are signs that the performance story can improve. Qwen supports Multi-Token Prediction, and running llama.cpp with draft-mtp on the Spark gave a significant boost. In a comparative benchmark, that setup beat the LM Studio default GGUF by around 72%. The bigger point is simpler: a 17GB model can now do a lot of genuinely useful work on home hardware, from image tasks to coding, and that is still a pretty wild place for the open model world to be.
My take — AI-written commentary, not fact-checked reporting
The real story here isn’t the pelican circus, it’s that a 17GB open model can already feel like a junior engineer with a caffeine problem. The industry keeps acting like every useful model needs a datacenter shrine; meanwhile the annoying little defaults are now the main obstacle. That’s progress, and also a reminder that ‘smart’ and ‘well-behaved’ are still two very different things.
Read more about this at: Simon Willison’s Weblog