Hermes adapter solves a SWE-Bench Pro task in 35 tool calls where the base Qwen3.8 model failed

yb2698 · x · 2026-11-19

Following NousResearch's Hermes adapter release for Alibaba's Qwen3.8-27B, the author shows a real SWE-Bench Pro (hard 51) task: the base model broke the test file and burned its 100-step budget, while the adapted model fixed the bug, verified task scenarios, and passed in just 35 tool calls. The adapter lifts SWE-Bench Pro pass@1 from 29.4% to 43.1% (51.0% with xhigh thinking) and Terminal Bench 2.1 from 44.7% to 58.8%, matching base-model xhigh performance with thinking off at 3.5–4.4× fewer tokens.

Original post →

More from coding & agent

coding & agent channel →