Qwen 3.8 27B: A Powerful but Over-Thinking AI Model

Alibaba’s Qwen lab has released Qwen 3.8 27B, a new Apache 2 licensed, vision-capable large language model. While it shows benchmark improvements over its predecessors and can produce impressive outputs, its default configuration presents a significant practical hurdle. The model is set by default to an ‘extra high’ reasoning effort, leading it to over-engineer solutions to simple prompts.

Key takeaways

• Qwen 3.8 27B is a new 27-billion-parameter, open-weight AI model from Alibaba with vision capabilities.
• Its default ‘xhigh’ reasoning setting causes it to over-think, using excessive time and tokens for simple tasks.
• Simple prompts like ‘draw an SVG of a circle’ can trigger minutes of planning for elaborate, animated outputs.
• The strong recommendation is to ignore the default and run the model on ‘low’ or no reasoning for practical use.
• Despite this issue, the model excels at complex tasks like generating detailed SVGs and performing vision analysis.

Impressive Performance with a Problematic Default

Qwen 3.8 27B is a new Apache 2 licensed, 27-billion-parameter vision-capable LLM from Alibaba’s Qwen lab. Its self-reported benchmarks show improvements over its predecessor, Qwen 3.6 27B, and the closed-weight Qwen 3.7-Plus.

However, the model’s default ‘reasoning_effort’ setting is ‘xhigh’ (extra high), which is described as a “hilarious” and bad default, especially for consumer hardware. This setting causes the model to over-think simple tasks, using excessive reasoning tokens and time. For example, when asked to generate an SVG of a pelican on a bicycle, the model took 21 minutes, using 22,276 reasoning tokens to produce 3,223 tokens of output. With a simpler prompt to “draw an svg of a circle,” the model spent several minutes planning an elaborate animated geometric study instead of producing a simple circle.

Recommendation for Users

The strong recommendation is to ignore the default. Users should run Qwen 3.8 27B on low or even no reasoning levels initially. While the model is capable of impressive outputs, the default ‘xhigh’ setting is not a practical starting point for most applications.

The Consequences of Excessive Reasoning

With the model’s default reasoning effort set to ‘xhigh’, it consumes large context windows, initially hitting LM Studio’s default 8,192 token limit for mundane problems. The documentation describes this as the default for complex tasks demanding thorough analysis, but it results in spectacular over-thinking on consumer hardware.

Over-Engineered Outputs

Even a simple prompt like “draw an svg of a circle” triggers lengthy reasoning. The model aims to create “a carefully crafted piece” with animation, layered rings, and a distinctive palette, producing an absolutely beautiful animated circle that was entirely not what was asked for. For a more complex “pelican riding a bicycle” SVG, the default setting used 22,276 reasoning tokens over 21 minutes to generate the output.

The Recommended Approach

The strong recommendation is to ignore that default and run Qwen 3.8 27B on ‘low’ or even no reasoning levels initially for faster, more appropriate responses. A test with reasoning turned off completed the pelican SVG prompt in 137 seconds—just over two minutes—compared to the 21 minutes required with the default ‘xhigh’ setting.

Capabilities and Practical Applications

Despite its tendency to over-think by default, Qwen 3.8 27B excels at specific, complex tasks. It can generate high-quality SVG images locally, as demonstrated by a detailed illustration of a pelican riding a bicycle. The model produced a well-structured SVG with a correctly shaped bicycle frame, properly positioned legs, a clear pelican pouch, wings touching the handlebars, and a tasteful background with sun, clouds, and hills.

Vision and Tool-Building Proficiency

The model performs well on vision tasks, such as accurately returning JSON bounding boxes for objects in a photograph. When prompted to identify pelicans on a 0-1000 scale, it correctly output coordinates that matched the image. Furthermore, Qwen 3.8 27B can build custom tools, like an interface for visualizing these bounding boxes. However, if the reasoning effort is not manually adjusted from its default high setting, it may over-engineer such solutions.

Its 27-billion-parameter size makes it suitable for running on reasonably specced laptops. A quantized build is available, with a 17GB file size noted for use in local testing.