Qwen 3.8 27B Shows Promise But Defaults to Excessive Analysis

Alibaba’s Qwen research lab recently released Qwen 3.8 27B, a vision-capable large language model (LLM) that’s generating considerable excitement in the AI community.

The model boasts impressive benchmark results, exceeding both its predecessor Qwen 3.6 27B and even Alibaba’s own closed-weight Qwen 3.7-Plus—a significant achievement considering the latter was one of their strongest models to date.

I’ve been testing this model on various hardware setups, including a MacBook Pro with 128GB M5 Max memory and an NVIDIA DGX Spark supercomputing platform.

The Overthinking Issue

The most notable characteristic I’ve observed is Qwen 3.8’s default tendency to overanalyze even simple prompts—a direct consequence of its configuration set to ‘xhigh’ reasoning effort by default.

This setting, while intended for complex tasks requiring in-depth analysis, consumes excessive resources and dramatically increases generation times. For instance, generating a relatively straightforward SVG image of a pelican riding a bicycle required 21 minutes and used 22,276 reasoning tokens to produce just 3,223 tokens of output.

By contrast, the same prompt with reasoning disabled generated in under two minutes using significantly fewer resources.

Impressive Capabilities Despite the Drawback

When properly configured, Qwen 3.8 demonstrates remarkable capabilities. The pelican-on-bicycle SVG I generated (with appropriate settings) showcased:

  • Correct bicycle frame shape
  • Legs positioned on both sides of the bike
  • A clearly defined pelican pouch
  • Wings extending to touch the handlebars—a rare feat in AI-generated images
  • Tasteful background with sun, clouds, and natural elements

These details highlight Qwen 3.8’s potential for creating high-quality visual content when not constrained by its default overthinking tendency.

The model represents a significant step forward in accessible LLMs, offering impressive performance in a relatively compact package (just 17GB on disk) that can run locally on consumer hardware.