HF releases 300MB RL-trained LoRA matching Qwen3.8-27B xhigh accuracy with 8.1× fewer tokens
Disastrous-Work-1632 · reddit · 2026-11-18
Aritra from Hugging Face released the second harness adapter: an RL-trained LoRA (r=32, 300MB) on a frozen Qwen3.8-27B, built inside the OpenCode coding agent.
- Key result: with thinking off, the adapter hits 51.0% on SWE-bench Pro (hard 51), matching the base model's xhigh-thinking accuracy (47.1%) with 8.1× fewer generated tokens (3.5k vs 28k median); 54.9% with thinking on.
- Run locally: full commands provided — hf download the Q80 GGUF and F16 LoRA, then llama serve with enablethinking: false and 64k context.
- Second in HF's harness-adapter series after pi, exploring small adapters tuned to a specific agent harness as a cheaper alternative to long thinking.
More from coding & agent
- New Hermes plugin adaption: fine-tune a model just by asking in chat — sarahookr · 2026-11-18
- Hermes agent can now fine-tune models in-chat via Adaption's AutoScientist plugin — sarahookr · 2026-11-18
- Sidehand MCP server drives Windows apps in the background at 2s vs 65s rivals — Agitated_Chair_4977 · 2026-11-18
- UK retailer M&S uses GitHub Copilot multi-agent pipeline to scale retail engineering — marlene_zw · 2026-11-18
- Voice agent team cuts P95 time-to-first-voice by 50% via single-datacenter migration — itsOmSarraf_ · 2026-11-18
- Pro tip: raise reasoning effort to stop GPT-6.2 Sol quitting early on long-horizon tasks — pvncher · 2026-11-18