Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things

Alibaba's Qwen 3.8 27B, a 27 billion parameter vision-capable language model, shows improved benchmarks over previous versions but defaults to an excessively high reasoning effort setting, causing slow outputs. Testing revealed that increasing context length improved performance, though generation times remain long for complex tasks.