Apple Silicon build for large models

This build suits someone who wants to run large models at home without the noise, heat, and power draw of a multi-GPU rig, and who is willing to accept slower generation in exchange.

Who it is for

Unified memory means most of the machine’s RAM is available to the model, so a high-memory Mac Studio can hold models that a single 24GB card cannot. That is the whole case for it. If your priority is the fastest possible tokens per second on mid-sized models, a discrete GPU will serve you better for the money.

Choosing the machine

Two current options sit at different points. The Mac Studio M4 Max offers up to 128GB of unified memory at a lower price, while the Mac Studio M3 Ultra has the highest memory bandwidth of any Apple Silicon, so it generates faster, though Apple narrowed its memory options during 2026. If buying used, an older M2 Ultra Studio can be good value. Confirm the current line-up and memory configurations before you buy.

What to expect

• Model sizes: comfortably runs large models that need more than 24GB, at reasonable quantisations.
• Speed: generation is slower than a high-end NVIDIA card, and prompt processing on long contexts is noticeably slower still. This is the main trade-off, and it is a real one.
• Living with it: near-silent and low-power, which makes it pleasant to have running on a desk.

Memory is fixed at purchase and cannot be upgraded later, so decide up front how much headroom you want. Buying more memory than you need today is the only way to leave room for larger models tomorrow.