Qwen 3.8 27B: strong open model, wildly overthinks by default

Alibaba's Qwen team shipped a 27-billion-parameter open model on Friday that looks genuinely useful for local work, then buried the useful part behind a default setting that makes it look terrible. Simon Willison's write-up of Qwen 3.8 27B is worth reading because the headline of his post, that the model 'defaults to wildly overthinking things,' is most of the story about where open reasoning models sit right now.

The specifics stick. Running the model at its default `xhigh` reasoning effort, Willison asked for the standard pelican-on-a-bicycle SVG and got a response that took nearly 21 minutes and used 22,276 reasoning tokens to produce 3,223 tokens of output. He calls the setting 'a hilarious default' and 'absolutely not a good way to run the model, especially on consumer hardware,' and his advice is blunt: turn reasoning down to low, or off, before you form an opinion.

Underneath the default, the model is the interesting part. It's Apache 2 licensed, 17GB as a quantized GGUF, vision-capable, with a 262,144-token context window, and Willison reports it drove a coding agent loop and did strong bounding-box detection on photos. Throughput on his local hardware, a 128GB MacBook Pro and an NVIDIA DGX Spark, landed at 15 to 30 tokens per second, well below the 74 to 184 tokens per second he sees from hosted APIs, though a Multi-Token Prediction path reportedly delivered around a 72% speedup over standard GGUF serving.

The post doesn't explain why Alibaba shipped `xhigh` as the default in the first place, or whether the self-reported benchmarks the team published were run at that setting. Both matter if you're deciding whether a headline win transfers to your own workload. Three researchers we track have already shared the write-up, a decent signal that the local-model community is paying close attention despite the rough edges.

For developers building on-device coding agents or vision tools, the release is a real step forward: an Apache 2 model that fits on a laptop and holds up in agent loops is not a small thing. The lesson for anyone shipping reasoning models is smaller and sharper. A default that costs users 21 minutes on their first prompt will earn one-star first impressions no matter how good the model actually is.