Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things

Simon Willison’s Weblog · 2 months ago · 44 ● 2 sources

Qwen Research released Qwen 3.8 27B, and tests showed the model’s default “extra high” reasoning can consume the full context and cause very slow, off-target outputs. The author’s first attempt with the full 262,144 max context length took 21 minutes (22,276 reasoning tokens) while producing 3,223 output tokens. Running with reasoning lowered or context reduced changes results to be faster and more task-faithful, and the model’s vision and bounding-box output work well enough for offline tooling even if the default setting tends to overthink.