27B local model with custom agent runtime builds a Word-like app from one instruction

bryany97 · reddit · 2026-10-07

Developer bryany97 released Aura, an open-source local cognitive system exploring whether architecture around a small model can meaningfully expand autonomous task capability.

Test prompt was a single line: clean-room reconstruct Microsoft Word, make it launchable, prove it works by writing a letter and exporting it
Aura queried its local Wikipedia corpus, generated its own behavioral spec, and wrote 44 checks before implementation so an empty program would fail
It then implemented 22 feature groups, launched the resulting app "Quill," wrote the proof letter inside it, and exported the PDF
Entire run used 27B local inference on a Mac, no frontier/cloud model

Repo and YouTube demo are public. The open question: how far can persistent agent runtimes push small local models?

Original post →

27B local model with custom agent runtime builds a Word-like app from one instruction

More from coding & agent