Qwen 3.8 27B is excellent, but default settings induce excessive reasoning

Released by Alibaba's Qwen team, Qwen 3.8 27B is a vision-capable LLM under the Apache 2 license. The 27B parameter size is appropriate for running on typical laptops, and it shows improved benchmark performance over the previous model, Qwen 3.6 27B.

However, this model has a default reasoning_effort setting of xhigh. While designed to perform thorough analysis for complex tasks, it induces excessive reasoning even for general tasks, potentially leading to inefficiencies in cost and speed.

In an actual SVG generation test, creating an image of 'a pelican riding a bicycle' took 21 minutes with the xhigh setting, but was completed in about 2 minutes with reasoning disabled. While the xhigh setting produces very sophisticated results, it consumes too many tokens and slows down speed in typical user environments.