Simon Willison on Qwen 3.8 27B: 22,276 Reasoning Tokens for 3,223 Output Tokens Because the Model Ships Defaulted to 'xhigh' Effort
Reviewing Alibaba's Apache-2.0, vision-capable Qwen 3.8 27B (17GB quantized, runnable locally), Simon Willison found the model 'defaults to wildly overthinking things' — its pelican-on-a-bicycle SVG took 21 minutes, burning 22,276 reasoning tokens to emit 3,223 output tokens, and a request for a plain circle triggered ruminations on design aesthetics. Disabling reasoning cut that to 137 seconds at visibly lower quality; LM Studio throughput ran 15–30 tok/s with Multi-Token Prediction adding roughly 72%. His recommendation is a one-line config fix rather than a verdict on the model: override the factory reasoning default, then it is genuinely usable for coding, vision, and tool-use. HuggingFace corroborates the traction — Qwen/Qwen3.8-27B at 415k downloads and 10.5k likes, with unsloth's GGUF at 2.73M downloads.
↳ Follow the thread