Alibaba's Qwen 3.8 27B launched August 16, 2026 as an Apache 2 licensed, vision-capable 27B parameter model. At 17GB quantized, it fits on consumer laptops and beats Qwen 3.6 27B and closed-weight Qwen 3.7-Plus on benchmarks.
The model defaults to xhigh reasoning effort, causing massive overthinking. A simple pelican SVG prompt consumed 22,276 reasoning tokens and 21 minutes to produce 3,223 output tokens. With reasoning disabled, the same task completed in 137s. Even "draw a circle" triggered minutes of architectural deliberation before generating an animated geometric study.
Vision capabilities remain strong: bounding box detection on pelican photos produced precise 0-1000 scale coordinates. The model also built a functional HTML labeling tool from a single prompt, though it over-engineered a demo scene with procedurally drawn pelicans.
Coding agent viability remains an open question. The model supports long context (262,144 tokens) and tool-calling, but the default reasoning setting makes it impractical for routine tasks. Recommendation: start with low or no reasoning.
Source: Hacker News · Summarized by HeadlinesBriefing