Qwen 3.8 27B ships Apache 2, but defaults to spectacular overthinking
Alibaba's Qwen 3.8 27B is a vision-capable, Apache 2 licensed model sized for laptop inference, but it defaults to xhigh reasoning effort that burns through context and takes 21 minutes to draw a pelican on a bike. Crank the reasoning_effort down and the same output drops to two minutes. A sharp reminder to check defaults before trusting a model's cost.