AnalysisAI ModelsAugust 17, 2026

Qwen 3.8 27B defaults to overthinking, says Simon Willison

Simon Willison tested Qwen 3.8 27B, an Apache 2 licensed 27B vision-capable LLM from Alibaba, on his M5 Max MacBook Pro and NVIDIA DGX Spark. He found its default xhigh reasoning effort causes spectacular overthinking, using 22,276 reasoning tokens to generate a simple SVG.

1 source

More stories today

Open the live feed