Researchers have benchmarked the performance of local language models running directly on mobile phones. The study measures critical metrics such as output speed, latency, and memory usage to evaluate the practical viability of pocket-scale AI inference.