Qwen3-0.6B on a 2017 Samsung Note 8 drives a real desktop Chrome session

A 400 MB local model running on a 2017 Samsung Galaxy Note 8 drove a real desktop Chrome browser through a relay, according to a developer who built the page-perception layer used in the test. The setup paired llama.cpp in Termux on the phone (Android 9, 6 GB RAM) with Qwen3-0.6B Q4_K_M, while a laptop ran Chrome in a normal, non-headless window.

The model never sees HTML, a screenshot, or a URL. It receives a structured representation of the page, roughly 200 tokens describing about 10 named links or fields, picks one by name, and copies facts it was given into JSON. The surrounding stack handles page capture, candidate selection, the click, reading facts, and verifying results.

The author tested 12 small models across three verifiable tasks with logs and offline replay. One task moved from an unrelated site to the Wikipedia Galaxy Note series page, where the model had to pick "Note 8" from similar names among roughly 760 interactive nodes and return the release date from the infobox. The author disclosed their role in building the perception layer and said the result proved more interesting than the original test.