Qwen3.8-27B's Hermes adapter trims the token tab in harness-hub's tests
Harness-hub reports higher coding and terminal scores than base Qwen3.8-27B under Hermes Agent, with thinking off.
TLDR
According to harness-hub's model card, the Qwen3.8-27B adapter lifts thinking-off scores from 29.4% to 43.1% on SWE-bench Pro's hardest 51 tasks and 44.7% to 58.8% on Terminal-Bench 2.1. It reports roughly matching the base model's xhigh-thinking accuracy with 3.5β4.4Γ fewer generated tokens. The catch, per the card: many thinking-on runs exhausted the context window in this Hermes Agent setup.
Combined views
7.8K
5 Sources, first seen ago

Comments