Qwen3-27B jumps from 39% to 65% on SWE-bench Pro with RL and one small LoRA

TheTuringPost · x · 2026-11-21

TheTuringPost recaps a Hugging Face recipe that flips the usual approach: keep the harness, adapt the model. With RL and a small LoRA, Qwen3.8-27B went from 39% to 65% on SWE-bench Pro. In a reply thread, rasbt notes LLMs are specifically fine-tuned for particular harnesses — the harness is part of the training setup, so the model can't absorb it, like a CPU/GPU absorbing the OS. The open question: which side is cheaper to change for your setup?

Related event: Debate: Will LLMs Absorb the Agent Harness, or Should Models Be Fine-Tuned(3 posts)→

Original post →

More from coding & agent

coding & agent channel →