Within 10k tokens, Hermes adapter solves 51% of TB-2.1 vs base's 28%

yb2698 · x · 2026-11-19

Within a 10k-token per-task budget, the Hermes adapter (thinking off) solves 51% of Terminal-Bench 2.1 and 39% of SWE-Bench Pro, while the base model with xhigh thinking manages only 28% and 0%. On tasks both solve, median tool calls drop from 23 to 7 on TB-2.1 and from 63 to 28 on SWE-Bench Pro.

Related event: NousResearch Releases Hermes Adapter for Qwen3.8-27B, Boosting SWE-Bench Pro to 43.1%(10 posts)→

Original post →

More from coding & agent

coding & agent channel →