Self-hosted Qwen3.8-27B + pi LoRA matches Opus 5.5 on a real bug fix in ~93s
yb2698 · x · 2026-11-18
A developer compared a self-hosted open model against a frontier model as a coding agent on a real bug: Qwen3.8-27B + pi LoRA (4×A6000, TP4) vs Opus 5.5, fixing lost request logs after logrotate in the same inference proxy. Key results:
- Both passed 60/60 tests; each solution passed the other's 9/9 tests
- Local model: 93s wall time, 95k input / 4.6k output tokens; human author took 1.5 min
- Same approach: a Recorder class doing one os.stat per write, reopening when the file is replaced or missing
- Notably, the local model ran its own tests and simplified one unprompted
Author's verdict: "Local AI will win."
More from coding & agent
- Early access to Devin bots and a chat with one of their creators, Walden Yan — Madisonkanna · 2026-11-18
- Open-sourced genetic algorithm optimizer hit top 2 in Open Kernel Challenge within 7 hours — morgymcg · 2026-11-18
- Future codebases will be 80% AI-assisted, raising sloppier-code concerns — 996roma · 2026-11-18
- Two weeks of service_tier "priority" A/B tests: median +5.5% TTFT, basically noise — iambubu0523 · 2026-11-18
- Sentry founder jokes about MCP as HN thread declares it the standard for AI agent tooling — zeeg · 2026-11-18
- Astra 6.2 hailed as best computer-use model after flawless MS Paint demo — gabrielchua · 2026-11-18