A 400MB model on a 2017 phone drives real Chrome

A developer built a page-perception layer for browser agents and tested 12 small local models against it. The smallest one won: a 400MB Qwen3-0.6B model, run on a Samsung Note 8 from 2017, drove a real desktop Chrome session rather than a simulated one, according to a post in r/LocalLLaMA.

If that holds, it resets what hardware a browser agent needs. Teams building these tools have been assuming the model needs real compute behind it, a cloud API or at least a recent phone. A nine-year-old midrange device driving an actual browser instead of a mock environment would mean the hardware floor for this kind of agent is a phone most people threw out years ago.

That's a big if. The post comes from one person with a product to build on top of this result, posted to a single subreddit thread, with no benchmark numbers, no video, no repo link, and no third party confirming the setup. "Drives real Chrome" is doing a lot of work in that sentence. The post doesn't say what tasks it completed, how reliably, or how it compares to running the same model on newer hardware.

Watch for someone outside that thread to replicate it, ideally with a task list and a failure rate. A model that clicks one button is a different claim than one that fills out a form or navigates a multi-step flow. Right now this is one builder's anecdote about their own tool, not an independent test.