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

Alibaba's Qwen research lab has unveiled Qwen 3.8 27B, a vision-capable LLM with 27 billion parameters released on Friday. This model builds upon the success of its predecessor, the 27B parameter Qwen 3.6, and outperforms Qwen 3.7-Plus, one of their strongest models until May 2024. Benchmarks indicate a significant improvement over its predecessors, but it defaults to an overthinking mode that can hinder performance on consumer hardware.

Despite this, Qwen 3.8 offers adjustable reasoning depth through the reasoning_effort parameter, allowing users to control cost and speed according to their needs. The model's impressive capabilities were demonstrated when it generated a pelican riding a bicycle SVG in 21 minutes, using 22,276 reasoning tokens to produce 3,223 tokens of output.

The model's ability to scale context length to 262,144 tokens allowed for this complex drawing, showcasing its potential in local machine execution.