Announcement
Qwen3.8-27B’s Hermes adapter cuts the token bill, its model card reports
The model card reports SWE-bench Pro hard-51 pass rate rising from 29.4% to 43.1%, with thinking off.
TLDR
Harness-hub’s model card reports that its Qwen3.8-27B adapter approaches the base model’s maximum-reasoning accuracy under Hermes Agent, with thinking off and 3.5–4.4× fewer generated tokens. It reports 43.1% on SWE-bench Pro’s 51 hardest tasks and 58.8% on Terminal-Bench 2.1. The catch: the card says many thinking-on runs exhausted the context window, so these results aren’t a blanket verdict on reasoning.
Combined views
9.9K
2 Sources, first seen ago

Comments