ExpertCache squeezes a 120B GPT model onto a 16 GB M1 Pro
ExpertCache lets a 120‑billion‑parameter GPT‑OSS model (63 GB) run on an M1 Pro with only 16 GB of RAM by paging expert tensors directly into Metal. The demo hits ~0.75 tok/s for prompts—far from interactive speed—but proves the checkpoint can stay functional without swapping. It opens a path to local inference on commodity Macs, albeit with performance trade‑offs.