Qwen 3.8 27B is excellent, but it defaults to overthinking things (simonwillison.net)

Article summary

The Qwen 3.8 27B model is a 27B parameter vision-capable LLM that defaults to overthinking, leading to slow performance. The model's default setting of xhigh for reasoning effort can result in extremely detailed and elaborate responses, but also increases processing time. The author suggests that adjusting the reasoning effort to medium or low can improve performance without significantly sacrificing accuracy. The model's ability to generate code, drive tools, and annotate images makes it a promising tool for various applications.

Main themes

• Qwen 3.8 27B model
• Overthinking and performance
• Reasoning effort settings
• Code generation and tooling
• Image annotation and vision capabilities
• Model optimization and fine-tuning

What commenters say

• The Qwen 3.8 27B model's tendency to overthink can be mitigated by adjusting the reasoning effort setting to medium or low, which can improve performance without sacrificing accuracy.
• Some users find the model's default setting of xhigh for reasoning effort to be excessive and unnecessary, leading to slow performance and wasted processing time.
• The model's ability to generate code and drive tools is impressive, but its performance is hindered by its tendency to overthink and its limited memory bandwidth.
• Fine-tuning the model with techniques such as LORAs or ThinkingCap can help reduce overthinking and improve performance, but may also affect the model's output quality.
• The trade-off between thinking time and output quality is a significant consideration for users, with some preferring faster but less detailed responses and others prioritizing accuracy and detail.
• The Qwen 3.8 27B model's performance can be improved with optimized hardware, such as GPUs with high memory bandwidth, but this may not be feasible for all users.
• Some users suggest using alternative models, such as Muse Glimmer, which may offer better performance and more efficient thinking patterns.
• The use of proxies or harnesses to control the model's thinking process and prevent overthinking can be an effective solution, but may require additional development and fine-tuning.