Qwen 3.8 27B defaults turn simple requests into 20-minute research projects

Alibaba’s Qwen lab released Qwen 3.8 27B on Friday, a 27-billion-parameter vision model under the Apache 2 license. The size sits at the sweet spot for running locally on a typical laptop, and its predecessor Qwen 3.6 27B has already proven itself. According to the company’s own benchmark tables, the new version outperforms both Qwen 3.6 and the closed-source Qwen 3.7-Plus, which was their strongest model of any size until May. Independent tests have not been published yet.

The model that runs on a laptop and thinks too much

The evaluation used two machines: a MacBook Pro M5 Max with 128 GB of RAM, and an NVIDIA DGX Spark. In both cases the model ran through LM Studio with a quantized build labeled Q4_K_M that occupies 17 GB on disk. Additional trials were run directly with llama-server on the Spark. The relatively compact file makes local execution feasible without a server farm, but the model’s default setting turns that advantage into a drawback.

Local hardware and 17 GB quantization

Official documentation sets the logical effort parameter (reasoning_effort) to xhigh by default, a level intended for complex tasks that require deep analysis. In practice this makes the model spend tens of thousands of tokens “thinking” even on trivial prompts. LM Studio limits the default to 8,192 context tokens, and the model fills the entire window immediately. Only after raising the limit to the maximum of 262,144 tokens does the issue disappear.

xhigh default produces extreme overthinking

When asked to “draw an SVG of a pelican riding a bicycle”, the model generated 22,276 reasoning tokens and 3,223 output tokens, taking 21 minutes. The resulting SVG was the best the reviewer had seen from a local model: correct bike frame, legs on both sides, a distinct beak pouch, wings touching the handlebars, motion lines behind, and a detailed background. The same request with reasoning disabled finished in 137 seconds and produced 3,715 tokens. The quality gap did not justify the wait.

21 minutes to draw a pelican on a bicycle

Even a minimal prompt like “draw an SVG of a circle” triggered a full “geometric circle research” chain: concentric circles, compass marks, graded fill, a subtle animation of a rotating wavy ring, and handling of prefers-reduced-motion—all for a single basic SVG element. The larger flagship model, Qwen 3.8 2.4T-A95B, released a week earlier, was accessed via OpenRouter and returned a simple animated SVG without this drama.