Alibaba’s Qwen lab released a second Qwen 3.8 model last week, and this one fits on a laptop. Where the 2.4 trillion parameter Qwen 3.8 flagship that AI Insiders covered on 15 August needs datacenter class hardware, Qwen 3.8 27B is a 27 billion parameter, vision capable model released under an Apache 2 license, small enough to run on a well specced consumer machine. Simon Willison, publishing on his own blog on 16 August, spent days testing it and reached a verdict that cuts two ways: the model is genuinely capable, and its default settings actively work against it.
Willison ran the 17GB Q4_K_M quantized build in LM Studio on a 128GB MacBook Pro and an Nvidia DGX Spark. Qwen’s own documentation sets the model’s reasoning_effort to xhigh by default, described there as intended “for complex tasks demanding thorough analysis.” Willison called that default “hilarious” and said it is “absolutely not a good way to run the model, especially on consumer hardware.”
The cost of that default is concrete, not theoretical. Willison’s pelican-on-a-bicycle SVG test took 21 minutes and burned 22,276 reasoning tokens to produce 3,223 tokens of usable output. With reasoning turned off, the identical prompt finished in 137 seconds. Asked for something as plain as an SVG of a circle, the model at its default setting spent several minutes reasoning through options like a “Bauhaus/compass-drawing vibe” color palette before returning an elaborate animated composition rather than the simple shape requested. Willison’s own conclusion: “My strong recommendation: ignore that default. Run Qwen 3.8 27B on low or even no reasoning levels at first.”
The overthinking is not purely wasteful, which is what makes the criticism worth carrying in full rather than softening into a shrug. On a harder task, building a bounding-box labeling tool from a single prompt, Willison found the model produced a working interface in one shot with reasoning on, while the same prompt with reasoning off returned a tool that placed the boxes in the wrong spot. He also put Qwen 3.8 27B behind a coding agent, Pi, inside a real codebase, and called the resulting output “very solid.” The model’s value tracks the task: trivial requests do not need xhigh, and Willison’s testing suggests genuinely hard ones might.
Speed is the model’s remaining weak point. Willison measured 15 to 30 tokens per second from LM Studio on both machines, well behind hosted alternatives: Artificial Analysis lists OpenAI’s 5.6 Sol at 74 tokens per second and 5.6 Luna at 184. Running the model with Multi-Token Prediction, a speculative-decoding trick built into the weights, produced roughly a 72 percent throughput gain in Willison’s own comparative benchmark, though getting there means bypassing LM Studio’s default server setup entirely.
Willison’s assessment is one experienced practitioner’s hands-on read across two machines, not a controlled evaluation against independent benchmarks, and he says so himself: he is still waiting to see what independent numbers show. That caveat limits how far the finding generalizes. It does not undercut the finding itself. The failure mode he documented, tokens and minutes spent deliberating over palette choices for a circle, is reproducible by anyone who downloads the same 17GB file and leaves the defaults untouched.
Any team evaluating Qwen 3.8 27B for local or self-hosted deployment should check the reasoning_effort default before wiring it into a production loop. Left at xhigh, it will multiply latency and token spend on requests that never needed deep reasoning in the first place, and the fix is a single configuration flag that a default-trusting integration could easily miss.
Simon Willison detailed his hands-on testing of Qwen 3.8 27B on his own blog, simonwillison.net, on 16 August 2026.
