Simon Willison reviews Qwen 3.8 27B, a 27-billion-parameter, vision-capable, Apache 2 licensed model from Alibaba's Qwen lab, successor to the well-regarded Qwen 3.6 27B. Qwen's own benchmarks claim it beats both Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus, though Willison notes independent verification is still pending.

He ran the 17GB Q4_K_M quantized build in LM Studio on a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark, plus llama-server directly on the Spark. The model defaults to an "xhigh" reasoning effort, which Willison calls a poor default for consumer hardware. Generating an SVG of a pelican on a bicycle took 21 minutes and 22,276 reasoning tokens for just 3,223 output tokens; with reasoning disabled, the same prompt finished in 137 seconds. A simple "draw an svg of a circle" prompt spiraled into a multi-minute meditation on Bauhaus color palettes and animated rings.

On vision tasks the model excelled: asked to return 0-1000 scale bounding boxes for pelicans in a photo, it nailed the coordinates. Willison then had the model build, in a single prompt, a full HTML tool to visualize such bounding boxes, though with reasoning left on it over-engineered the result, adding an unrequested demo scene with hand-drawn pelican silhouettes.

As a coding agent, paired with the lightweight Pi tool and served via LM Studio over Tailscale, Qwen 3.8 27B successfully explained the authentication logic in Willison's Datasette codebase and wrote a working Python script to convert a JSONL session transcript to Markdown.

Speed remains the main weakness: 15-30 tokens per second on his hardware, well behind hosted models like OpenAI's 5.6 Sol (74 tokens/second) and 5.6 Luna (184 tokens/second) per Artificial Analysis figures. Following a tip from llama.cpp creator Georgi Gerganov, Willison tested the model with Multi-Token Prediction enabled via llama.cpp's draft-mtp mode, which delivered roughly a 72% speed improvement over the default LM Studio setup in a GPT-5.6-run benchmark.

His takeaway: a 17GB file now delivers coding, vision and tool-calling capability that would have required expensive datacenter hardware a year ago, and the main obstacle to daily use is the poorly chosen default reasoning setting rather than the model's underlying capability.