snapbench

How fast do local LLMs run on Snapdragon phones and laptops, and which llama.cpp backend wins? Every result runs the same pinned llama.cpp build, the same models and quants, and records which GPU driver was active. Numbers are tokens per second: pp512 is reading a 512-token prompt, tg128 is generating 128 tokens. Higher is better; the fastest backend in each row is highlighted.