Qwen 3.8 27B matches trillion-parameter models on a laptop

Alibaba's Qwen team released Qwen 3.8 27B, and the interesting part is the size. The model scores 52 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Luna and sitting one point behind GLM-5.2 and DeepSeek V4 Pro 0813. Those two carry 753 billion and 1.7 trillion parameters. Qwen 3.8 27B has 27 billion. Simon Willison, who ran it on his own machines, calls it "a truly astonishing model" for that reason: near-frontier quality that fits on a laptop.

It is Apache 2 licensed and vision-capable, and a quantized build is about 17GB, small enough to run on a 128GB MacBook Pro or an NVIDIA DGX Spark with a context window up to 262,144 tokens. The catch is that it overthinks. By default it runs at "xhigh" reasoning effort, so asking it to "draw an svg of a circle" sends it off designing a geometric study with concentric guide circles and animation. Willison's pelican-on-a-bicycle test took 21 minutes and 22,276 reasoning tokens. With reasoning turned off, the same prompt finished in 137 seconds. The reasoning_effort parameter controls this, and for most work you will want it set well below the default.

Why it matters

If you run models locally, a 27B open-weight model that trades blows with trillion-parameter systems changes what you can do on your own hardware, but plan to tune reasoning_effort down before you trust it in an agent loop, or a one-line prompt can burn twenty minutes.