Pi-Adapted Qwen3.8-27B Doubles Local Coding Pass Rate With 63% Fewer Tool Calls
yb2698 · x · 2026-11-19
Hugging Face adapted Qwen3.8-27B for the pi coding agent with RL and a LoRA: SWE-bench Pro (Hard 51) pass@1 jumped 29.4% → 62.7%, Terminal-Bench 2.1 39.5% → 55.8%, with 55–62% fewer tool calls; weights and methodology are open. User kesslerIO then ran the merged Q4KM quant locally on a 128 GB M5 Max MacBook Pro with Splash: 796 tok/s prefill and 119 tok/s decode at 19.5 GiB on a 32K prompt, completing 13/24 coding attempts vs 7/24 for base Q4 with 63% fewer tool calls (425 vs 1141). He notes paired reliability p=0.125, so general superiority isn't established.
More from coding & agent
- Dirac Robotics Launches CAD-to-URDF, But a Skeptic Says It Costs a Few Euros of Tokens to DIY — kamathsblog · 2026-11-19
- Dev laments copy-pasting between GPT-6.2 Astra and Fable 5.2: 'I'm a meat proxy' — Angaisb_ · 2026-11-19
- Swyx: Build a Business Where New Model Releases Make You Happier, Not Obsolete — js_craft_hq · 2026-11-19
- Dev's ideal multi-model stack: plan with Fable, execute with Sol, review with Astra — jdjohnson · 2026-11-19
- Google launches Gemini 4 Neon: #1 on AA Coding Agent Index, 50-80% cheaper per task — VeryWellVersed · 2026-11-19
- Open-source playbook catalogs 117 App Store rejection patterns, gains 542 stars in a day — mjmirza · 2026-11-19