MLX Peer: splitting local inference across a Mac and iPhone

Field notes

The page describes a layer-partitioned inference pass: the Mac handles embeddings and output, while the iPhone runs layers 0–11 and returns activations over USB. Its exporter writes separate weight files without first allocating the full model in memory. In the cited 27B run, the phone held 2.57 GB of weights and decode reached 3.93 tokens per second after the first token; the page flags substantial Mac swap and a failed prefill tolerance check.