One rank-96 LoRA boosts Qwen3.8-27B across 3 agent harnesses, 55.5% to 63.5% mean
HauntingMoment · reddit · 2026-11-20
A "super adapter" experiment for Qwen3.8-27B: LoRAs trained separately for three agent harnesses (Hermes Agent, OpenCode, pi) turned out nearly orthogonal (cosine similarity 0.01), so they were concatenated along the rank dimension into a single rank-96 LoRA—no lossy merging (no TIES/DARE).
Results (thinking on, xhigh):
- SWE-Bench Pro (hard 51): Hermes Agent 45%→61%, OpenCode 47%→67%, pi flat at 74%.
- Terminal-Bench 2.1: +4 to +6 points across all three harnesses.
- Mean over 6 harness/benchmark pairs: 55.5%→63.5%.
- Efficiency (tasks both solve): 53-73% fewer tool calls with thinking off, 19-48% with thinking on.
The blog's core is a reusable recipe: train a harness adapter for your own model, harness, and tasks. Adapter released at harness-hub/Qwen3.8-27B-super-adapter; full experiments and open questions in the HF blog post.
More from coding & agent
- Build a 'company brain': one shared knowledge base for your whole team's AI — VibeMarketer_ · 2026-11-21
- LangChain ships four updates: LangSmith Agent Platform, trace data controls, Tool Gateway — LangChain · 2026-11-21
- Stanford CS329Z: Full AI Agents Course Free With Slides, Videos, Homework — stanfordnlp · 2026-11-21
- Grok Bot Gets Its Own Email Addresses to Sign Up and Handle Promos for You — elonmusk · 2026-11-21
- CapCut × ChatGPT plugin turns chat into a video generator and editor — SimplyAnnisa · 2026-11-21
- Princeton & Meta's Strategist evolves inference strategies, beating baselines on 24/30 benchmarks at 39% lower cost — 141_1337 · 2026-11-21