Agentic data science with the Data Agent Kit
cloud.google.com·8h ago
Data Science
Simon Willison's Blog·Aug 16·5 min read
Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things
TL;DR
Qwen 3.8 27B, a new vision-capable language model from Alibaba, shows significant performance improvements but tends to overthink responses by default.
✦ Why It Matters
Experiment with Qwen 3.8 27B's settings to optimize its response quality for your specific use case.
Key Takeaways
- Run Qwen 3.8 27B with low or no reasoning for faster results.
- Utilize Multi-Token Prediction to enhance model performance.
- Test the model on various hardware to find optimal settings.
- Explore independent benchmarks to validate Qwen's performance claims.
How It Works
Qwen 3.8 27B employs a reasoning_effort parameter that adjusts the depth of analysis during output generation. The default setting, xhigh, leads to thorough but often unnecessary reasoning, resulting in longer processing times.
Users can switch to medium or low settings for a more efficient experience.