Qwen 3.8 27B is excellent, but it defaults to overthinking things
Simon Willison tests Alibaba's Qwen 3.8 27B, a 17GB local LLM that impresses with vision, coding, and bounding box tasks, but its default 'xhigh' reasoning effort makes it overthink even simple prompts, leading to slow, over-engineered outputs. He recommends turning down the reasoning for most tasks.
This is a hilarious default. It's absolutely not a good way to run the model, especially on consumer hardware.