Nous Hermes adapter lifts Qwen3.8-27B SWE-Bench Pro from 29.4% to 43.1% with 3.5-4.4x fewer tokens
yb2698 · x · 2026-11-19
Nous Research released a Hermes adapter for Alibaba's Qwen3.8-27B targeting agentic coding efficiency:
- SWE-Bench Pro (hard 51) pass@1: 29.4% → 43.1%; with xhigh thinking: 45.1% → 51.0%
- Terminal-Bench 2.1 pass@1: 44.7% → 58.8%
- Key selling point: with thinking off, it matches or beats the base model's xhigh thinking performance while using 3.5-4.4x fewer tokens
For open-weight agentic coding workflows, this means near-flagship performance at a fraction of the inference cost.
Related event: Nous Releases RL-Trained Hermes Adapter for Qwen3.8-27B(13 posts)→
More from coding & agent
- Garry Tan's "Personal AGI" is a folder of markdown files — 220,000 pages, 400x his 2013 output — jasonkneen · 2026-11-19
- Stanford Professors Launch Noeri, Self-Improving AI That Companies Own Instead of Rent — YejinChoinka · 2026-11-19
- Same weights, 3x score: Gemini 4 Neon jumps 12.8%→39% on SWE-sweep by swapping the harness — Delicious_Character6 · 2026-11-19
- ESP-Skainet: Open-Source Framework for Fully Offline Voice AI on ESP32 — alexcovo_eth · 2026-11-19
- Ploten ships an MCP server that live-edits open Power BI Desktop reports — plotenhq · 2026-11-19
- Testing gemini-4-neon with a DIY PowerShell agent harness: "benchmarks are for tourists" — dfinke · 2026-11-19