Hermes adapter gets GGUF release: matches xhigh thinking at a quarter of the tokens

yb2698 · x · 2026-11-19

harness-hub released a GGUF build of the Qwen3.8-27B Hermes LoRA adapter for llama.cpp, trained with TRL AsyncGRPOTrainer plus custom RL envs, OpenEnv and safeagents. With thinking off, the adapter matches the base model at xhigh reasoning effort while generating 3.5–4.4× fewer tokens per task — a big win for local deployment. The GGUF file itself hasn't been re-evaluated through llama.cpp, and a quantized base would shift results further.

Original post →

More from coding & agent

coding & agent channel →