Summary & Key Takeaways ​

• Qwen 3.8 27B is a new Apache 2 licensed, vision-capable LLM from Alibaba. • It's an excellent size for running on reasonably specced laptops. • Benchmarks from Qwen show significant improvements over previous versions. • The model defaults to an "xhigh" reasoning effort, leading to extensive internal thought processes. • This default causes the model to "wildly overthink" even simple prompts. • Overthinking results in high token usage and significantly longer generation times. • Increasing the context length can mitigate issues caused by the default reasoning effort.

Our Commentary ​

This is such a classic LLM quirk, and Simon nails it. We've seen models that are too eager to please, but 'wildly overthinking' is a new level of hilarious frustration. It's a crucial heads-up for anyone looking to deploy Qwen 3.8 locally; that default setting could burn through context windows and compute cycles faster than you can say 'pelican riding a bicycle.' It makes me wonder how many other models have similar hidden 'personality traits' that aren't immediately obvious from benchmarks alone. This kind of practical, hands-on testing is invaluable.