A 300MB adapter lifts a local 27B coding model from 35% to 51% on SWE-bench Pro, beating max reasoning with thinking off
First seen on X 2 days ago@TeksEdge ♥ 1,61715 views
A post about a 300MB add-on that makes a local 27B coding model much better drew 1,600 likes and 2,100 bookmarks on X. It's Qwen3.8-27B-opencode-adapted, released Nov. 18 by the Harness Hub team on Hugging Face.
According to the model card, it's a LoRA adapter for Alibaba's Qwen3.8-27B, trained with reinforcement learning to run behind the open-source coding agent OpenCode. On a hard 51-task subset of SWE-bench Pro it goes from 35.3% to 51.0% (+15.7 points), and on Terminal-Bench 2.1 from 46.4% to 52.4%. On tasks both versions solve, median turns drop from 48 to 13. With thinking off, the adapter beats the base model at its maximum reasoning effort on SWE-bench Pro (51.0% vs 47.1%) while generating six to eight times fewer tokens, the team says.
It's Apache 2.0 and ships with instructions for vLLM, transformers and llama.cpp, plus GGUF files, so it runs on a home GPU. The card recommends leaving thinking off since turning it on adds nothing significant, and the team says it will publish the full training recipe so anyone can build adapters for other harnesses.
The caveat: every number comes from the OpenCode harness with one attempt per task and a 100-step budget, so the team says it isn't comparable to the official Qwen card, which uses other harnesses such as Claude Code and the full SWE-bench Pro. One reply said weight adapters "make your local models of any type dramatically better."
Comments 0
Select a sentence in the story (long-press on phones) to comment on it.
“So now a small 300MB adapter makes your local 27B coding model dramatically better!”
