Apple M4 Max SoC

The headline number on the M4 Max isn't the CPU core count — it's the memory bandwidth and the fact that the GPU and CPU address one shared pool. That is what makes it a serious local-inference machine rather than just a fast laptop chip.

Why it matters for LLMs

On a discrete-GPU setup, model weights have to be copied across the PCIe bus into VRAM before the GPU can touch them, and anything that doesn't fit spills to system RAM at a brutal speed penalty. On the M4 Max, the weights, the KV cache and the working buffer all live in the same unified memory — no copy, no PCIe ceiling.

Practical setup

• Run models through Ollama or LM Studio — both use Apple's Metal backend and pull GGUF quants directly.
• Size the model to leave headroom: macOS and your other apps need several gigabytes, so on a 64 GB machine treat ~48 GB as your realistic weight-plus-cache budget.
• A Q4_K_M quant of a 70B model fits comfortably on the 64 GB and 128 GB configurations; the hardware sizer gives you the exact figure for a specific model.

The rule

Buy the memory, not the core count. The gap between a 36 GB and a 128 GB M4 Max is the difference between running 8–13B models and running 70B-class models locally.