Run Open Models on Your Laptop Without Guessing RAM

Most people pick a local model the wrong way.

They ask which model is smartest. The better question is whether the model, quantization, context cache, and runtime fit in the memory you actually have. Get that wrong and a "70B on my Mac" demo becomes a swap-thrashing mess. Get it right and a 7B–8B Q4 instruct model is already useful for private chat, coding help, document Q&A, and low-latency experiments — with nothing leaving your machine.

This is the setup I wish I had on one page.

(This post continues for paid subscribers on Substack.)