Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things
Friday's big release was Qwen 3.8 27B , an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba's Qwen research lab. I've been looking forward to this one: 27B is an excellent size for running a model on a reasonably specced laptop, and its predecessor Qwen 3.6 27B was impressive. Qwen's self-reported benchmarks for this model are eye-opening. They show a boost from both Qwen 3.6 27B…
Qwen 3.8 27B, a vision-capable LLM from Alibaba's Qwen research lab, has recently been released. This 27B parameter model is excellent for running on a reasonably specced laptop, and its predecessor, Qwen 3.6 27B, was also impressive. The model's self-reported benchmarks show a significant improvement over previous versions, such as Qwen 3.6 27B and Qwen 3.7-Plus.
I have been running the model on two different machines: a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark. Despite Qwen's documentation describing the model as defaulting to xhigh for reasoning effort, which is not ideal for consumer hardware, I found the results to be entertaining. The default context limit of 8,192 tokens in LM Studio proved problematic, but increasing the context length to 262,144 eliminated the issue.
I tested the model's ability to draw bounding boxes around pelicans in a photograph, and it managed to produce a full interface from a single prompt, even drawing its own pelicans purely because of the label "pelicans" used in the example JSON. Additionally, I generated a 1000x1000 placeholder image with two blob-like "pelican" silhouettes placed at the given bboxes, showcasing the boxes aligning with the pelican image.
Written by urgent.news from Simon Willison's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
