Engineer Clusters Old Android Phones to Run a 27B LLM Locally

An engineer successfully clustered two old Android phones using a custom C++ RPC engine and Hadamard ternary quantization to run the 27B Ternary-Bonsai model locally. This hardware-split workaround successfully bypassed local memory limits and achieved a stable speed of ~2.8 tokens per second. This project demonstrates impressive systems-level engineering in edge AI, proving that resource-constrained mobile devices can run massive models without thermal throttling. It offers a creative pathway for developers looking to bypass traditional RAM and VRAM limitations on consumer hardware. The setup uses ternary weights (-1, 0, 1) folded in GGUF with a custom patch that applies the Hadamard transform at runtime on the ARM CPU, replacing heavy MAC operations with additions and subtractions. Consequently, power consumption remains low, keeping maximum device temperatures around 30°C.