Qwen 3.8 27B: Default Settings Cause Overthinking Problem

Alibaba’s Qwen research lab released the 27B-parameter vision-capable large language model “Qwen 3.8 27B” under the Apache 2 license. While it demonstrates high performance, it has drawn attention for a problem where default reasoning settings lead to extremely inefficient results. According to Simon Willison’s blog post, the model adopts “xhigh” reasoning effort as its default, and when running on consumer-grade hardware, output generation takes an astonishing amount of time due to verbose thinking processes.

High Performance Demonstrated in Benchmarks

Qwen 3.8 27B’s self-evaluation benchmarks show results surpassing its predecessor, Qwen 3.6 27B, and the unreleased Qwen 3.7-Plus. The fact that it outperforms Qwen 3.7-Plus, one of the most powerful Qwen models as of May 2026, is technically intriguing because it was achieved at the relatively small parameter scale of 27B. Simon Willison’s blog post expresses the view that independent benchmark results are awaited.

The 27B parameter size makes the model suitable for local execution on relatively high-spec laptops. Expectations for this release were high because the predecessor model, Qwen 3.6 27B, had produced impressive results. Indeed, Willison verified the model on two different machines: a MacBook Pro (128GB M5 Max) and an NVIDIA DGX Spark.

The “Histrionic Overthinking”

Caused by Default Settings

The official documentation for Qwen 3.8 27B describes features for controlling reasoning effort. Three levels are available: xhigh (default), medium, and low. The default xhigh is described as being intended for “complex tasks requiring thorough analysis.” However, Willison describes this default setting as “histrionic and by no means an appropriate way to run on consumer hardware.”

As a specific example of the problem, with LM Studio’s default context length (8,192 tokens), even considering everyday tasks consumes the entire context. This issue was resolved by expanding to the model’s maximum context length of 262,144 tokens. However, reasoning in xhigh mode remains extremely compute-intensive.

In an actual test, an SVG image generation task of a pelican riding a bicycle was presented. In xhigh mode, it required 21 minutes of processing, consuming 22,276 reasoning tokens before eventually generating 3,223 output tokens. The resulting SVG was evaluated as being of extremely high quality, including the pelican’s shape, the bicycle frame, leg positioning, wing placement, and background composition.

Performance Comparison with Reasoning Disabled

When run with the reasoning feature disabled using the same prompt, output generation took only 137 seconds (approximately 2 minutes 17 seconds) and generated 3,715 tokens. Compared to the 21 minutes in xhigh mode, the time required is reduced to about one-ninth. Willison states plainly: “Was it worth waiting 21 minutes? Absolutely not.”

Furthermore, a comparison was also conducted using the larger Qwen 3.8 2.4T-A95B model via OpenRouter. The results suggest that local execution of the 27B model could become a practical alternative under appropriate reasoning settings.

The Importance of Configuring Settings

According to Use Case

This case demonstrates that verifying default settings is essential in the operation of large language models. In particular, parameters that control the depth of “reasoning effort” and “chain-of-thought” are directly linked to the tradeoff between cost and quality. It is speculated that the decision to make xhigh the default reflects an intention to guarantee the highest quality output for all potential tasks, but real-world operation requires staged setting adjustments according to task complexity.

The fact that a 27B-parameter model exists on disk as a 17GB quantized file and runs in a local environment expands the possibilities for use in edge computing and highly privacy-constrained environments. However, the reality that these advantages cannot be fully realized without optimizing reasoning settings has also been brought into sharp relief.

Editorial Opinion

In the short term, the arrival of Qwen 3.8 27B brings a new option to the local LLM market. Its 27B parameter size allows it to run even on laptops with 16GB–32GB of memory, providing engineers and researchers with an accessible test environment. However, the default setting issue significantly complicates cost calculations for adoption. Over the next three to six months, local LLM runtime environments, starting with LM Studio, are expected to actively offer reasoning effort presets for Qwen 3.8. In the long term, dynamic control of reasoning effort may become established as a standard feature of large language models. Development of mechanisms that assess task complexity in advance and automatically adjust reasoning effort will likely accelerate. Furthermore, Alibaba’s top-tier public benchmark results with a 27B model demonstrate that Chinese AI research institutions maintain their technical competitiveness, which will influence future model development competition. The fundamental question raised by this case is whether the burden of balancing “maximizing the model’s capabilities” and “operating at a practical cost” should be borne by developers or users.

References

• “Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things”, by Simon Willison — Simon Willison’s Weblog, 2026-08-16T22:00:39.000Z (ARR)
• Source URL: https://simonwillison.net/2026/Aug/16/qwen-38-27b/