300MB RL-trained LoRA lifts local Qwen3.8-27B coding agent 15.7% on SWE-bench Pro
yb2698 · x · 2026-11-19
Hugging Face released an RL-trained LoRA adapter (300MB, Apache 2.0) for Qwen3.8-27B, optimized for the OpenCode coding harness and runnable locally via llama.cpp. Reported results:
- SWE-bench Pro (51 hard tasks): 35.3% → 51.0% (+15.7 pts)
- Terminal-Bench 2.1: 46.4% → 52.4%
- Median agent turns on shared solves: 48 → 13 (73% fewer turns)
- With thinking disabled, the adapted model approaches or exceeds the base model at max reasoning effort while generating 6–8× fewer tokens per task
A cheap, local way to substantially speed up coding agents.
Related event: HF's 300MB RL LoRA lifts Qwen3.8-27B to 51% on SWE-bench Pro locally(5 posts)→
More from coding & agent
- To Try Gemini 4 Neon, Blogger Jumps Through Antigravity Signup Hoops — vista8 · 2026-11-19
- Power user details his $200/mo Grok agent setup: chief-of-staff bot, 5 job applications in 7 minutes — elonmusk · 2026-11-19
- Pass-the-Device: Transplanting VPN/SASE Trust After Endpoint Compromise — cyb3rops · 2026-11-19
- BeeKeeper: Open-Source Pixel-Art Hive Visualizes Your Codex and Claude Code Agents — RanAviv · 2026-11-19
- Dimillian shares pro prompt tips for AI-generated games: PBR, post-processing are the cheat codes — Dimillian · 2026-11-19
- Solo dev launches Sokudo, an IDE to run Codex, Claude Code and other agent CLIs side by side — DeVoresyah · 2026-11-19