A super-adapter for the Qwen3.8-27B model raised the mean pass rate from 55.5% to 63.5% across six benchmark-harness pairs using xhigh thinking, according to data from developer yb2698. The adapter, which supports the Pi, OpenCode, and Hermes Agent harnesses, reduced tool call volume on tasks solved by both the base and adapted models by 19% to 48% with thinking enabled and by 53% to 73% with thinking disabled.
The lightweight LoRA adapter allows the model to match the performance of the base model’s highest reasoning level while using 3.5 to 4.4 times fewer tokens when thinking is off. The Hermes-adapted version of the model has been ported for local inference on Mac hardware using MLX builds.
