The models are memory mapped from disk so the kernel handles reading them into memory. As long as there's nothing else requesting that RAM, those pages remain c

The models are memory mapped from disk so the kernel handles reading them into memory. As long as there's nothing else requesting that RAM, those pages remain cached in memory between invocations of the command. On my 128 GB workstation, I can use several different 7B models on CPU and they all remain cached.