- Learn
- Pick your hardware
Running LLMs on a Mac: which M-series chip and how much memory
Why Macs with Apple silicon are good at local AI, how much unified memory you need for each model size, and how they compare to a PC with a GPU.
A Mac with Apple silicon is one of the easiest ways to run local AI, and with enough memory one of the few ways to run really big models in a quiet box on your desk.
Why Macs are good at this
On a PC, a model has to fit in the graphics card's memory, usually 8 to 24 GB. A Mac shares one pool of unified memory between the processor and the GPU, and you can buy it with 64, 128 GB or more. That fits models no single consumer graphics card can hold.
The trade-off is speed. Generating text is limited by memory bandwidth (see why bandwidth sets speed), and only the Max and Ultra chips have GPU-class bandwidth.
Chip matters as much as memory
| Chip | Bandwidth | Qwen3 8B, estimated |
|---|---|---|
| M2 | 100 GB/s | ~11 tok/s |
| M4 Pro | 273 GB/s | ~29 tok/s |
| M4 Max | 546 GB/s | ~58 tok/s |
| M3 Ultra | 819 GB/s | ~87 tok/s |
A base M-chip is fine for small models. For anything 30B and up, look at a Max or Ultra.
How much memory
macOS lets the GPU use about two thirds of memory on Macs up to 36 GB and about three quarters above that. Roughly:
| Unified memory | Usable for models | Runs well |
|---|---|---|
| 16 GB | ~11 GB | Qwen3 8B |
| 24 GB | ~16 GB | gpt-oss 20B, Qwen3 14B |
| 36 to 48 GB | ~24 to 36 GB | Qwen3 30B-A3B, Qwen3 32B, Gemma 3 27B |
| 64 GB | ~48 GB | Llama 3.3 70B (slowly) |
| 128 GB+ | ~96 GB+ | gpt-oss 120B |
If you know what you're doing, you can raise the GPU's share at your own risk, for example to 56 GB on a 64 GB Mac (it resets on reboot):
sudo sysctl iogpu.wired_limit_mb=57344The big-model sweet spot: mixture of experts
Big dense models like Llama 3.3 70B run at single-digit speeds on most Macs. Mixture-of-experts models are a much better fit, because they need lots of memory but read only a small part of it per token. gpt-oss 120B on a 128 GB M4 Max runs at an estimated 60+ tok/s, while Llama 3.3 70B manages about 7. See mixture of experts explained.
Mac or PC?
- Mac: silent, low power (about 100 W), simple, and the only easy way to 128 GB+. Slower per dollar for models that fit on a GPU, and much slower at reading long prompts.
- PC with a used 3090: faster for anything that fits in 24 GB, and cheaper. Loud, hot, and needs a second card for 70B models.
A used Mac Studio with 64 GB is our "Quiet" build on the builds page.
Software
Ollama works out of the box and uses the GPU automatically. LM Studio also supports MLX, Apple's own format, which is often a little faster on Macs. See Ollama vs llama.cpp vs LM Studio.
One email when there's a new guide, podcast episode, or result from the rack. No spam.