Qwen 3.8 27B is excellent, but it defaults to overthinking things

7.7 relevance

Qwen 3.8 27B model analysis with high community engagement, directly relevant to AI/ML.

General simonwillison.net

Answer the board on AI readiness

Arrive with evidence, not opinion. See how your foundation benchmarks against your peers, get a clear read, and leave with a ranked place to start.

Why It Matters

For developers running local LLMs on laptops or edge hardware, this highlights how default reasoning settings can cripple inference performance and token budgets, making it critical to tune reasoning_effort and context length for practical use.

Key Takeaways

Always override the default reasoning_effort to 'low' or 'medium' when running Qwen 3.8 27B locally to avoid excessive token consumption and generation times.

Qwen 3.8 27B is excellent, but it defaults to overthinking things

Summary

Alibaba's Qwen 3.8 27B, an Apache 2.0 licensed vision-capable LLM, defaults to an 'extra high' reasoning effort that consumes excessive tokens even for trivial prompts, generating elaborate but unintended outputs like an animated circle SVG instead of a simple static one. Running the 17GB Q4_K_M quantized build on a 128GB M5 Max MacBook Pro or NVIDIA DGX Spark, a pelican SVG took 21 minutes and 22,276 reasoning tokens to produce 3,223 output tokens, versus 137 seconds with reasoning disabled. The model supports adjustable reasoning_effort to control cost, but the default setting makes it impractical for consumer hardware without increasing context limits beyond LM Studio's default 8,192 tokens.

Author

Simon Willison