The phone tier is a deployment choice determined by the target surface, not the model. A phone SoC combines a CPU, GPU, and integrated NPU; its runtime and portability options differ from SBC and small-box deployments. section 11.6 section 11.4
Phone deployments use TFLite and LiteRT-LM or Core ML and MLX. Runtime choice matters: on Apple silicon, the compute-unit hint can outweigh weight-variant choice, and the runtime may silently override it. Measure end-to-end latency, since preprocessing and postprocessing can exceed inference time. section 11.6
Phone-tier devices
The Pixel 10 Pro and iPhone 16 are phone-tier examples. For a Pixel audio pipeline, a Kotlin log-mel STFT can take longer than the Whisper-tiny encoder it feeds; a NEON-tuned FFT can reduce more latency than an inference-side change. section 11.4 section 11.6
