According to a write-up by developer Simon Willison, Alibaba's Qwen research lab released Qwen 3.8 27B, an Apache 2 licensed, 27-billion-parameter vision-capable large language model. Willison, whose account is the single source for this report, tested a 17GB Q4_K_M quantized build via LM Studio on a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark. He noted the model defaults to an 'xhigh' reasoning-effort setting, which he described as 'not a good way to run the model, especially on consumer hardware.' In one test, generating an SVG of a pelican riding a bicycle took 21 minutes and used 22,276 reasoning tokens to produce 3,223 tokens of output. Running the same prompt with reasoning turned off produced output in 137 seconds. He also hit LM Studio's default 8,192-token context limit, which the model exhausted on mundane problems; loading the full 262,144 context length resolved it. Willison reported strong results on a bounding-box test around pelicans in a photo, calling the match 'such a good match.' He noted Qwen's self-reported benchmarks show gains over Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus, and said independent benchmarks are still awaited. Qwen supports a 'reasoning_effort' parameter to control reasoning depth and cost.
- 27B parameter, Apache 2 licensed, vision-capable model from Alibaba's Qwen lab
- 17GB Q4_K_M quant runs on a laptop; defaults to 'xhigh' reasoning effort
- One SVG prompt took 21 minutes with reasoning on vs 137 seconds with it off
- Supports a reasoning_effort parameter to control reasoning depth and cost
What it means for you
A capable, freely-licensed AI model that you can run on your own machine (no cloud, no subscription) is now available and works well. The catch: it ships with a setting that makes it 'overthink' every request, so a simple task can take 20 minutes instead of two. Turning that setting down fixes it.
Try this
If you experiment with local models in LM Studio, load Qwen 3.8 27B and set 'reasoning_effort' to a lower level (and raise the context length from the 8,192 default) before running anything — you'll get near-instant results instead of multi-minute waits.
Who should care
Tinkerers and small teams who want to run AI locally for privacy or cost reasons and have a reasonably specced machine (roughly 32GB+ RAM). Not a fit for people who just use ChatGPT or Claude in a browser.
Skip this if
You don't run models on your own hardware and have no intention of starting — this changes nothing about the tools you already use.
Sources: Simon Willison — read the original