Qwen 3.8 27B: Powerful but Overthinks by Default
Original: Qwen 3.8 27B is excellent, but it defaults to overthinking things
Why This Matters
A high-quality 27B open-weight model with vision capability expands practical local AI deployment options.
Alibaba's Qwen lab released Qwen 3.8 27B, an Apache 2-licensed 27B-parameter vision-capable LLM, on August 15, 2026. Blogger Simon Willison tested the model on an M5 Max MacBook Pro and NVIDIA DGX Spark, finding strong output quality undermined by an 'xhigh' default reasoning setting that causes extreme token overuse.
Qwen 3.8 27B is a 27-billion-parameter, Apache 2-licensed, vision-capable large language model released by Alibaba's Qwen research lab. Simon Willison tested it using LM Studio's 17GB Q4_K_M quantized build on both an M5 Max MacBook Pro (128GB) and an NVIDIA DGX Spark.
The model defaults to 'xhigh' reasoning effort, which Qwen's documentation describes as suited for 'complex tasks demanding thorough analysis.' In practice, Willison found this setting caused dramatic overuse of tokens even for trivial prompts. A simple request to draw an SVG of a circle triggered elaborate internal reasoning about adding concentric rings, animations, and gradient fills.
For a pelican-riding-a-bicycle SVG test, the model consumed 22,276 reasoning tokens and 3,223 output tokens, taking 21 minutes to complete. With reasoning disabled, the same prompt finished in approximately 137 seconds. Willison noted the xhigh output was visually the best local-model pelican SVG he had seen, but called the 21-minute wait 'absolutely not' worth it.
LM Studio's default 8,192-token context limit was quickly exhausted by the model's reasoning traces; Willison resolved this by setting the context window to the full 262,144 maximum. Qwen's supported reasoning_effort levels are xhigh (default), medium, and low. Willison's conclusion is that xhigh is a poor default for consumer hardware, and users should manually lower the setting for everyday tasks.