320MB LoRA RL-trained in Hermes Agent gives Qwen3.8-27B thinking-level accuracy with 3.5–4.4× fewer tokens
HauntingMoment · reddit · 2026-11-19
Nous Research RL-trained a small LoRA adapter (r=32, 320MB F16 GGUF) on a frozen Qwen3.8-27B inside Hermes Agent — their third harness adapter after pi and OpenCode.
Results (same tasks, one attempt each):
- SWE-bench Pro (hard 51): base 29.4% → adapter 43.1% with thinking off, nearing base-with-thinking's 45.1%; 51.0% with thinking on
- Terminal-Bench 2.1: 58.8% with thinking off, matching base with thinking (xhigh)
- With thinking off: 19 newly solved Terminal-Bench tasks vs 7 lost (p=0.03); 10 new solves vs 3 lost on SWE-bench Pro
The upshot: thinking-off accuracy nears thinking-level while using 3.5–4.4× fewer tokens. GGUF weights are released with a one-command llama.cpp setup (plus vLLM/transformers/PEFT merge notes), all eval trajectories are public, and the full recipe ships in a Friday blog post.
More from coding & agent
- Google's GUARD paper: frozen base model, self-editing agent harness gains +8.89 points — omarsar0 · 2026-11-19
- Microsoft open-sources openclaw-dev: one-command hosted OpenClaw agent on Azure with sandboxed tool execution — adnan_hashmi · 2026-11-19
- 6-minute Hubble documentary made with Claude and Jimeng CLI workflows — xiaohu · 2026-11-19
- Coding agents gone wild: test-cheating, scope drift, and rogue DB migrations — Common_Dream9420 · 2026-11-19
- Regenerative software: keeping intent as the source of truth so code can always be regenerated — gamagamer · 2026-11-19
- Every on Claude Fable 5.2: Human Thinking at the Start Matters More as AI Builds More — every · 2026-11-19