HF releases RL-trained Qwen3.8 adapter for opencode: +16pts on SWE-bench Pro, 8x fewer tokens
LysandreJik · x · 2026-11-18
HuggingFace's Lysandre Jik introduced Qwen3.8-27B-opencode-adapted, an open-weights LoRA adapter trained with RL specifically for the opencode coding agent, runnable locally. With thinking off, pass@1 jumps from 35.3% to 51.0% on SWE-bench Pro (Hard 51) and 47.0% to 53.0% on Terminal-Bench 2.1; with thinking on (xhigh) it reaches 54.9% and 56.6%. Notably, thinking-off matches base-model thinking accuracy while using just 1/8th the tokens on SWE-bench Pro and 1/6th on Terminal-Bench, and 12x less wall-clock time per task. Weights are open, with a writeup on the training method.
More from coding & agent
- Auto-approving PRs: Qodo's rules-based workflow handles the coding-agent PR bottleneck — omarsar0 · 2026-11-18
- Dev critique: GPT models' strict instruction-following kills productivity; astra gets stuck in self-blocking loops — kevinkern · 2026-11-18
- Record your judgment process 3 times, let AI build business evals from it — alliekmiller · 2026-11-18
- GPT-6.2 Sol hits 54.7 on Bug Hunt for $7, matching Fable 5.2 at 6% of the cost — PawelHuryn · 2026-11-18
- Noodle: Free open-source Mac app runs collaborative personal AI agents locally — Dave_Maynor · 2026-11-18
- Top engineers have stopped reading code: specs and tests become the persistent artifact — ZeeshanZiaML · 2026-11-18