Qwen 3.7 27B, running as a 17GB GGUF in LM Studio, drew Simon Willison the best pelican-on-a-bicycle he has seen from any local model. The demo signals on-device image generation has crossed a qualitative threshold.
Key facts
- Qwen 3.7 27B runs as 17GB GGUF file
- Tested in LM Studio on M5 Max laptop
- Simon Willison: best local-model pelican image
- 27B parameter class at Q4/Q5 quantization
- Official Qwen 3.7 release notes not yet published
Simon Willison, the well-known Python and open-source developer, posted on X that the new Qwen 3.7 27B — running locally as a 17GB GGUF file in LM Studio on his M5 Max laptop — produced the best pelican riding a bicycle he has seen from any model that runs on his hardware. According to @simonw, the image quality marks a notable step forward for locally-run image generation.
What the 17GB GGUF tells us
A 17GB GGUF for a 27B-parameter model implies aggressive quantization, likely Q4 or Q5 precision. That file size fits comfortably within the M5 Max's unified memory pool, which starts at 36GB in base configurations. The practical implication: users can run Qwen 3.7 27B alongside a browser, editor, and other tools without memory pressure. Willison's phrasing — "best I've seen from any model that runs on my laptop" — is a comparative claim against prior local models, not against cloud-hosted frontier systems.
Why the pelican test matters
Image-generation benchmarks often miss the qualitative dimension. A pelican riding a bicycle is a compositionally awkward subject — proportionally unusual, requiring coherent anatomy and a plausible mechanical interaction. That a 27B model running at consumer quantization handles it cleanly suggests the underlying vision-language alignment has improved beyond what the previous 3.0 27B delivered. The company has not yet published official release notes for Qwen 3.7, so details on training data, architecture changes, or benchmark scores remain undisclosed.
The local-model trajectory
Willison's demo fits a pattern from the past 90 days: mid-size models (7B-32B) are absorbing capabilities that were cloud-only a year ago. The M5 Max's memory bandwidth — publicly known to exceed previous Apple silicon generations — is the enabling hardware factor. As quantization tooling improves and GGUF support in LM Studio matures, the gap between local and hosted image generation narrows further. The pelican is a small data point, but it lands in a consistent trend line.
Key Takeaways
- Qwen 3.7 27B, as a 17GB GGUF, drew Simon Willison's best local pelican-bicycle image.
- Signals on-device generation crossed a quality threshold.
What to watch
Watch for the official Qwen 3.7 release notes and whether the company publishes benchmarks against the previous 3.0 27B. Also track whether LM Studio's GGUF support expands to other mid-size vision-language models in the coming weeks, which would confirm the local-generation trend is broad rather than Qwen-specific.
