Your Code Doesn't Have to Leave the Building
A local LLM is three layers: weights on disk, a runtime that loads them, an interface that talks to it. Once you see the stack that way, the hardware question answers itself and the tuning stops being guesswork.