Simon Willison tested Alibaba’s Qwen 3.8 27B, a 27-billion-parameter vision-capable open-weight model that fits in a 17GB file, and found its default xhigh reasoning setting caused extreme overthinking: a single pelican-riding-a-bicycle SVG generation took 21 minutes and 22,276 reasoning tokens versus about 2 minutes with reasoning disabled. He also tested Multi-Token Prediction support in llama.cpp, measuring a 72% speed improvement, and ran the model as a coding agent via the Pi framework, though token generation speeds of 15-30 tokens per second remained slower than hosted APIs.
