| |
Alibaba's Qwen 3.8 27B is a capable open-source vision language model that performs well on benchmarks, but defaults to "xhigh" reasoning effort, causing it to dramatically overthink even simple tasks and consume excessive tokens and processing time. When run locally on consumer hardware with default settings, the model used up all available context attempting mundane problems, though adjusting the reasoning effort to lower settings produces reasonable results in minutes rather than the 21 minutes the overthinking default requires. Despite this inefficient default, the model shows impressive capabilities for its 27B parameter size when properly configured.
Read Full Article →
← More Tech news