Qwen 3.8 27B is excellent, but it defaults to overthinking things
Points and comments are a snapshot, not live.
Qwen 3.8 27B is a capable local LLM but defaults to excessive reasoning effort.
Alibaba's Qwen 3.8 27B is an Apache 2 licensed, 27B parameter vision-capable LLM that runs on consumer hardware. Simon Willison tested it on a MacBook Pro and NVIDIA DGX Spark, finding its default `xhigh` reasoning effort leads to extremely long generation times and token waste (e.g., 21 minutes for an SVG). He recommends using `low` or no reasoning settings. The model demonstrates strong performance in bounding box detection and coding agent tasks when reasoning is dialed down. Speed remains a catch at 15-30 tokens/second, though Multi-Token Prediction can boost performance by ~72%.
What commenters are saying
Commenters widely agree the overthinking problem is severe and that the `xhigh` default is impractical. Many recommend using lower reasoning efforts or alternative models like Muse Glimmer, which is terser and more token-efficient. One user shares a proxy that cuts off thinking after 2K tokens and redirects the model. Another notes that applying a ThinkingCap Lora can reduce thinking tokens by ~40% without major quality loss. A few point out that the speed penalty on dense models like this is significant, making token efficiency a key operational cost for agents.