Alibaba's Qwen 3.8 27B LLM Ships with Overthinking Default
TL;DR. Alibaba's Qwen research lab released the new Apache 2 licensed Qwen 3.8 27B vision-capable LLM, an upgrade to its 27B parameter models. - The model's default reasoning effort is set to 'xhigh', causing it to overthink tasks, especially on consumer hardware. - This setting prioritizes thorough analysis over speed and cost, making it inefficient for practical deployment without adjustment. - Independent benchmarks are still awaited, but early user experience highlights the need to adjust reasoning parameters.
- Alibaba's Qwen 3.8 27B, a vision-capable LLM, is now available under Apache 2 license.
- The model ships with a 'xhigh' reasoning_effort default, leading to excessive processing for common tasks.
- Users report that the default setting makes the model inefficient on consumer hardware and requires manual adjustment.
- Qwen's internal benchmarks show significant improvement over prior versions, pending third-party validation.
Sources
- Qwen 3.8 27B is excellent, but it defaults to overthinking things — simonwillison.net
- artificialanalysis.ai — artificialanalysis.ai