Qwen 3.8 27B Hobbled by Its Default Overthinking Setting

👁 119 reads

Alibaba’s Qwen research lab released Qwen 3.8 27B on 15 August 2026 - a 27-billion-parameter, Apache 2 licensed, vision-capable model whose default reasoning setting causes it to massively over-generate tokens even on trivial prompts.

Developer Simon Willison tested the model on a 128 GB M5 Max MacBook Pro and an NVIDIA DGX Spark, in both cases using LM Studio with a 17 GB Q4_K_M quantized build. Qwen’s own benchmarks show the model outperforming both its predecessor Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus, which Willison notes was one of Qwen’s strongest models of any size as recently as May 2026.

The central problem is the out-of-the-box configuration. Qwen’s documentation lists four reasoning effort levels - xhigh, medium, low, and off - with xhigh as the default, described as intended “for complex tasks demanding thorough analysis.” In practice, the model applied that level to a request as simple as drawing a circle, with its internal reasoning trace beginning:

“The user is asking for an SVG drawing of a circle. Simple request - but I want it to be a carefully crafted piece.”

A prompt to draw a pelican riding a bicycle at xhigh consumed 22,276 reasoning tokens to produce 3,223 tokens of output and took 21 minutes to complete. The same prompt with reasoning disabled produced 3,715 tokens in 137 seconds. The xhigh run also immediately exhausted LM Studio’s default 8,192-token context window; loading the model with its full 262,144-token maximum context resolved that issue.

Despite the inefficiency, Willison praised the quality of the xhigh SVG output - noting correct bicycle frame shape, legs on both sides, a clear pelican pouch, wings touching the handlebars, and motion lines placed correctly behind the figure.

Practical takeaway: Anyone running Qwen 3.8 27B locally should override the default reasoning effort immediately. Set it to low, medium, or off for everyday tasks, and also raise the context limit beyond the 8,192-token LM Studio default - otherwise the model silently truncates mid-thought. Reserve xhigh for genuinely complex analytical work where quality justifies the wait.

// read moreQwen 3.8 27B Matches GPT-5.6 Luna on Intelligence IndexBenchmarksQwen 3.8 vs GPT-5.6 Luna: 7x Price Gap, Same ContextBenchmarksQwen 27B: Style Training Changed AI Behavior UnexpectedlyBenchmarks

Kinetyra