Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things
Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba's Qwen research lab, was released on Friday. Simon Willison tested it on his 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark, using LM Studio's 17GB Q4_K_M quantized build. He found the model excellent but noted a problematic default: the reasoning effort is set to 'xhigh', which leads to spectacular overthinking. For example, generating an SVG of a pelican riding a bicycle took 21 minutes, using 22,276 reasoning tokens to produce 3,223 output tokens. The same prompt with reasoning turned off took just over two minutes. Even a simple request to draw a circle triggered an elaborate reasoning trace and produced an animated circle. Willison recommends ignoring the default and running the model on low or no reasoning levels initially. He also tested the model's bounding box capabilities, which performed well, accurately returning bounding boxes for pelicans in a photo.
Developers running Qwen 3.8 27B locally should adjust the reasoning effort to avoid excessive token usage and slow generation.