News · 2026-08-16

Qwen3.8-27B spent 22,276 thinking tokens on one drawing

Alibaba's new open-weight Qwen3.8-27B ships configured to think as hard as possible on every request, and the cost of that default is measurable. Independent developer Simon Willison found that a single request to draw a pelican riding a bicycle took 21 minutes and burned 22,276 reasoning tokens to produce 3,223 tokens of actual output. The same prompt with reasoning switched off finished in 137 seconds.

Key facts

• The first attempt consumed roughly seven reasoning tokens for every token in the final answer.
• Willison published the measurements on August 16, 2026, the weekend after the model's Friday release.
• The model is an Apache 2.0 licensed, vision-capable 27B model from Alibaba's Qwen lab, with a 262,144-token context window.
• Primary source: Simon Willison's writeup and the Qwen3.8-27B model card.

The release itself was well received. Ground Truth covered the argument over whether Qwen3.8-27B is a relabelled predecessor earlier this week. This is the deployment story, and it is the one that will actually cost people money.

Qwen3.8-27B exposes an official setting called reasoning_effort with four levels. The documentation describes xhigh as being "for complex tasks demanding thorough analysis", medium as balancing accuracy and speed, and low as efficient reasoning optimized for speed and cost. Qwen sets xhigh as the default, and the widely used LM Studio build preserves it. Willison's verdict is direct: "This is a hilarious default. It's absolutely not a good way to run the model, especially on consumer hardware."

To understand why that matters, it helps to know what reasoning tokens are. Modern models are trained to write out a private working-through of a problem before producing an answer. Those intermediate tokens are generated at the same speed and, on hosted APIs, billed at the same rate as visible output, but the user never sees them. It is the difference between a colleague who answers your question and one who thinks aloud for twenty minutes first, at your expense, before saying the same thing.

The failure mode is easiest to see on a trivial prompt. Willison asked the model, at the default setting, to "draw an svg of a circle". The reasoning trace shows the model talking itself into a project: it considers concentric guide circles like a compass drawing, tick marks, a soft gradient fill, a slowly rotating dashed ring, a pulsing glow, and debates palettes ("deep teal ink on warm paper? Or bold vermilion circle on off-white with navy construction lines"). Several minutes later it produced an animated geometric study, which Willison notes was "entirely not what I had asked for".

There is a second, sharper trap for people running the model locally. LM Studio's default context limit is 8,192 tokens, and at xhigh the model was consuming the entire budget thinking about mundane problems before it could answer. Loading the model with its full 262,144-token context made the problem go away, but a user who never touches that setting will experience an expensive model that appears to fail on easy questions.

None of this makes the model bad. Willison calls the resulting pelican drawing "by far the best pelican SVG I've been able to generate with a model that runs on a local machine", from a file just 17GB on disk. He also found the model strong at returning bounding boxes around objects in photographs, a standard practical test for a vision model. Qwen's own published comparisons show gains over both the previous 27B and the closed-weight Qwen3.7-Plus. The problem is purely the shipped configuration.

His recommendation is unambiguous: "My strong recommendation: ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first."

The honest caveat is that this is one developer's measurement on two machines, using one quantized build, and the numbers will move with different hardware, quantization, and prompts. Willison himself notes he is waiting to see what independent benchmarks say about the model's quality. What the measurement does establish, and what Qwen's own documentation confirms, is that the default is xhigh and that the setting is doing exactly what it says.

The broader pattern is worth naming. The industry has spent two years learning that spending more compute at inference time buys accuracy on hard problems. What that research does not say is that every problem is hard. Shipping a model with maximum thinking enabled turns an optional accuracy tool into a mandatory tax, and on consumer hardware, where you are paying in minutes rather than cents, the tax is impossible to ignore.

For background, see our explainers on chain-of-thought reasoning, quantization, and open-weight models. The model itself is on Hugging Face, and LM Studio is the runner Willison used.

Primary source, verified: read the paper →

Ground Truth. (2026, August 16). Qwen3.8-27B spent 22,276 thinking tokens on one drawing. Ground Truth. https://groundtruth.day/news/qwen3-8-27b-spent-22276-thinking-tokens-on-one-drawing.html

@misc{groundtruth:qwen3-8-27b-spent-22276-thinking-tokens-on-one-drawing,
  title  = {Qwen3.8-27B spent 22,276 thinking tokens on one drawing},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/qwen3-8-27b-spent-22276-thinking-tokens-on-one-drawing.html}
}

Topics: qwen · local-models · open-weights · reasoning · inference-cost · china