Qwen 3.8 27B is strong on a laptop, but its default reasoning turns 137 seconds into 21 minutes
Simon Willison ran Alibaba's Apache 2 licensed Qwen 3.8 27B as a 17GB Q4_K_M build in LM Studio on a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark, praising its coding-agent work, bounding-box precision, and image annotation at 15-30 tokens per second. The shipped xhigh reasoning default is the trap: one pelican SVG burned 22,276 reasoning tokens over 21 minutes, against 137 seconds with reasoning off. Running it locally means starting at low or no reasoning and lifting the context past LM Studio's 8,192-token default, and speed — not output quality — is the remaining barrier to a 17GB local model replacing a hosted API.
Source: simonwillison.net ↗
It took 21 minutes to generate, using 22,276 reasoning tokens to produce 3,223 tokens of output.
Why this matters
- → Default reasoning setting makes 27B model 10× slower on local hardware
- → Strong coding and vision tasks hide a performance trap for consumer deployment
- → Context limits expose overthinking as a major UX barrier, not an output-quality win