Key Info

Nous Research announced a Hermes adapter for Alibaba's Qwen3.8-27B model, aimed at more efficient and accurate agentic development.

Highlights

  • SWE-Bench Pro (hard 51) pass@1 rises from 29.4% to 43.1%, and from 45.1% to 51.0% with xhigh thinking.
  • Terminal-Bench 2.1 pass@1 rises from 44.7% to 58.8%; with xhigh thinking it moves from 58.8% to 57.6%.
  • With thinking off, the adapter matches the base model's xhigh-thinking performance while using 3.5-4.4x fewer tokens.