Qwen 3.8 27B ships with its reasoning dial cranked to max

Alibaba's Qwen lab released Qwen 3.8 27B, an Apache 2 licensed vision-capable model that runs from a 17GB Q4_K_M quantized file. Simon Willison ran it on a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark and reports it can generate code, call tools, and annotate images — but that the shipped default reasoning effort of xhigh is a poor starting point.

The concrete failure mode is worth knowing before you download it. At the default setting, the model exhausted LM Studio's 8,192-token context thinking about trivial prompts; loading the full 262,144-token context fixed that. One SVG generation took 21 minutes and 22,276 reasoning tokens to produce 3,223 tokens of output. Asked simply to "draw an svg of a circle," it deliberated over palettes and animation and returned something other than what was asked. The same prompt with reasoning off finished in 137 seconds. Willison's recommendation is to start at low or with reasoning disabled entirely.

Reasoning isn't useless here: with it off, a one-shot bounding-box labelling tool rendered boxes in the wrong place. On vision, the model returned bounding boxes on a 0-1000 scale that matched two pelicans in a photo closely.

The open question is speed. Willison measured roughly 15-30 tokens/second under LM Studio and calls that the main barrier to daily use, citing dense non-MoE memory bandwidth demands. Running llama.cpp with --spec-type draft-mtp to use the model's Multi-Token Prediction support beat the LM Studio default by around 72% in a benchmark he had GPT-5.6 run. Independent benchmarks of Qwen's self-reported gains aren't in yet.