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

Alibaba's Qwen research lab has unveiled Qwen 3.8 27B, a vision-capable language model with 27 billion parameters, licensed under Apache 2. The model's impressive benchmarks showcase significant improvements over its predecessor, Qwen 3.6 27B, and a closed-weight model, Qwen 3.7-Plus. Running the model on two different machines, the author observed that it defaults to an overly cautious "xhigh" reasoning effort, which is not ideal for consumer hardware.

Despite this, the author found the model's large context length to be a valuable feature, allowing for detailed and creative outputs, such as generating an animated SVG of a pelican riding a bicycle. When the reasoning depth was reduced, the model produced a much faster and more concise output, demonstrating the trade-off between cost and accuracy.

The author acknowledged that while the default reasoning effort may be amusing, it is not practical for most use cases.