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

<p>Friday's big release was <a href="https://huggingface.co/Qwen/Qwen3.8-27B">Qwen 3.8 27B</a>, an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba's Qwen research lab. I've been looking forward to this one: 27B is an excellent size for running a model on a reasonably specced laptop, and its predecessor <a href="https://simonwillison.net/2026/Apr/22/qwen36-27b/">Qwen 3.6 27B</a> was impressive.</p> <p>Qwen's <a href="https://huggingface.co/Qwen/Qwen3.8-27B#benchmark-results">self-reported benchmarks</a> for this model are eye-opening. They show a boost from both Qwen 3.6 27B <em>and</em> the closed-weight Qwen 3.7-Plus, which was one of Qwen's strongest models of any size as recently as <a href="https://qwen.ai/blog?id=qwen3.7-plus">May this year</a>. It will be interesting to hear what independent benchmarks have to say about the model.</p> <p>I've been running the model on two different machines: my 128GB M5 Max MacBook Pro, and an <a href="https://simonwillison.net/2025/Oct/14/nvidia-dgx-spark/">NVIDIA DGX Spark</a>. On both machines I'm running LM Studio and <a href="https://lmstudio.ai/models/qwen3.8">their 17GB Q4_K_M quantized build</a>. I also tried using <code>llama-server</code> directly on the Spark.</p> <h4 id="the-default-of-extra-high-results-in-spectacular-over-thinking">The default of extra high results in spectacular over-thinking</h4> <p>Qwen's documentation describes the model as defaulting to <code>xhigh</code> for the reasoning effort, and the LM Studio GGUF I've been trying preserves that default:</p> <blockquote> <p>Qwen3.8 comes with official support for <code>reasoning_effort</code>, which can be used to adjust reasoning depth and control cost:</p> <ul> <li> <code>xhigh</code> (default): for complex tasks demanding thorough analysis</li> <li> <code>medium</code>: balancing accuracy and speed</li> <li> <code>low</code>: efficient reasoning optimizing for speed and cost</li> </ul> </blockquote> <p>This is a <em>hilarious</em> d