Fully local pipeline: open-weight 27B models ship production code with Playwright + vision-loop tests
awitod · reddit · 2026-11-21
A Redditor demos an all-local AI pipeline built on an RTX 5090/4090 workstation and Strix Halo: a 12-minute video (days of runtime, hundreds of millions of tokens) where even the talking-head avatar is AI—ComfyUI + Wan video + InfiniteTalk, Ming3 voice cloning, Qwen ASR, and a Qwen 27B (QFN quant). The substance: using open-weight 27B models to make complex changes to a real platform, gated by CodeQL multi-language analysis and Playwright UI integration tests, with a vision loop that catches UI bugs—thousands of tests passing, zero build warnings. Toolchain: llama.cpp, gufo, audio.cpp, ffmpeg, ComfyUI. Evidence that fully local open models can run a production dev-test loop.
More from coding & agent
- OpenAI details when Ultrafast speed on GPT-6.2 is worth it for rapid iteration — OpenAIDevs · 2026-11-21
- Agent swarm builds ice-flow PDE solver, discovers glacier sliding law beating textbook models — DynamicWebPaige · 2026-11-21
- Paradigm CTO: be unreasonably ambitious with AI — rewrite your whole dependency stack in days — nptacek · 2026-11-21
- Tip: Add Antigravity to your phone's home screen for remote control of all agent threads — amankhan · 2026-11-21
- The sentences your customer-facing AI agent must never say—and how to test for them — rcleffke · 2026-11-21
- Haijun Xia pushes back on 'UI is dead': chat is a UI too, and far from sufficient — sh_reya · 2026-11-21