Local models can now power computer use agents, but regulated industries still lack a playbook

Ambitious_Fold_2874 · reddit · 2026-10-01

The author built a computer use agent harness via Claude Code, powered locally by qwen3.8 flash next nvfp4: drawing a simple image took 12m 1s, largely spent debugging a WebGL issue; local speeds ran 1000 tps prefill and 50 tps generation. They note OpenAI DevDay's ultrafast cloud models (like "astra ultrafast") work faster for a premium.

They ask for open-source, plug-and-play computer use harnesses robust enough for regulated fields like law and medicine, plus experiences on HIPAA compliance, ensuring workflow accuracy, and choosing local vs cloud models.

Original post →

More from coding & agent

coding & agent channel →