Qwen3-0.6B on a 2017 Samsung Galaxy Note 8 Automates Desktop Chrome Browser Tasks
Developers demonstrated that a 400 MB quantized Qwen3-0.6B model running locally on a 2017 Samsung Galaxy Note 8 via llama.cpp inside Termux can successfully automate tasks on a desktop Chrome browser. Using a custom structured page-perception layer, the tiny model achieved a 100% success rate on multiple test tasks without ever processing raw HTML or screenshots. This demonstration proves that agentic browser automation does not require massive cloud LLMs or modern hardware when paired with an effective abstraction layer that reduces token overhead. It highlights the potential for legacy mobile phones and low-power edge devices to perform complex, automated web workflows locally and privately. The structured page representation reduced prompt size from hundreds of thousands of raw HTML characters down to about 200–500 tokens, enabling task completion in under 80 seconds. However, the author noted clear limitations: while sub-1B models excelled at name matching and data extraction, reasoning tasks requiring complex page judgment caused failures in models under 1.5B parameters.
## BACKGROUND
Running Large Language Models (LLMs) on resource-constrained edge devices typically relies on quantization techniques like Q4_K_M to compress model weights while retaining accuracy. Tools such as llama.cpp enable efficient local C/C++ model execution on CPU architectures, while Termux provides a Linux terminal emulator environment on Android. Traditional web agent frameworks feed raw HTML or page screenshots directly into large vision-language models, consuming massive amounts of memory and token context.