Qwen3.8-27B LoRA lifts SWE-bench Pro 35.3%→51%; tweet claims $ANT drops 10%
yb2698 · x · 2026-11-19
harness-hub released a GGUF LoRA adapter trained with RL inside OpenCode on Qwen3.8-27B:
- SWE-bench Pro (hard 51): 35.3% → 51.0% (+15.7 points)
- Terminal-Bench 2.1: 46.4% → 52.4% (+6.0; +13.6 on tasks not used for checkpoint picking, p=0.035)
- Uses 1/4 the turns on SWE-bench Pro and 1/2 on Terminal-Bench 2.1
- With thinking off it matches the base model's max-effort (xhigh) accuracy using 6–8× fewer tokens
- Adapter-only file is 319 MB on top of the base GGUF, Apache-2.0
The poster claims the release knocked $ANT down 10% and asks why pay for Claude — unverified.
More from coding & agent
- CoHDL turns PCB schematics into typed source code that compiles to netlists, built for AI — juntao · 2026-11-19
- Watch an AI agent make real phone calls: full demo with live transcripts and human handoff — mattyp · 2026-11-19
- Hands-on coding test: Tencent's Hy4 Flash beats DeepSeek by 12x on same task — HeyToha · 2026-11-19
- LangChain's Max Agency podcast returns for season 2 with Clay's Head of AI — LangChain · 2026-11-19
- Garry Tan's "Personal AGI" is a folder of markdown files — 220,000 pages, 400x his 2013 output — jasonkneen · 2026-11-19
- Same weights, 3x score: Gemini 4 Neon jumps 12.8%→39% on SWE-sweep by swapping the harness — Delicious_Character6 · 2026-11-19