Comparing Qwen3.8-27B fine-tunes and baselining vs. frontier

A Reddit post compares Qwen3.8-27B fine-tunes against a baseline and frontier models, detailing performance metrics such as median tokens and instances of models exceeding a 16,384-token limit. Specifically, Signal-Terse-Coder had 1 violation, while Dirk and Unsloth each had 4. The author developed a tool for this analysis, now available on GitHub at https://github.com/ashe-wb/tuieval, designed to be extensible for other users.

This report uniquely offers a direct comparison of Qwen3.8-27B fine-tunes against both baseline and frontier models, unlike most analyses that focus on a single model type.