A Mac’s processor and GPU share one pool of memory (unified memory), so a Mac with a lot of memory can hold models that no consumer graphics card can. Three things to know.
1. Not all of it is available to the GPU
macOS limits how much unified memory the GPU may use. Reports for recent Apple silicon give roughly two-thirds on smaller-memory machines and about three-quarters on larger ones. Our estimates plan around 67% up to 36 GB of memory and 75% above that, and keep a little in reserve. A 16 GB Mac is realistically a 10 to 11 GB model machine.
2. You can raise the limit, with care
On recent versions of macOS you can raise the cap with a Terminal command, for example sudo sysctl iogpu.wired_limit_mb=<megabytes>. The setting does not survive a restart. Leave several gigabytes for macOS and your other apps: if the system runs short of memory it will slow to a crawl or become unstable. The setting name and behavior have changed between macOS versions, so check current instructions for yours before using it. We do not recommend it unless a model misses by a small margin.
3. Use MLX when you can
MLX is Apple’s array framework, built for this hardware. MLX versions of models (formats) are often faster than GGUF on a Mac. LM Studio runs both; MLX LM is the command-line route. Ollama uses GGUF.
What to expect
- Speed follows the chip’s memory bandwidth, which differs a lot between base, Pro, Max and Ultra chips (why). Check the figure for your exact chip on its hardware page.
- Long prompts take a while to start compared with a fast graphics card, because reading the prompt is limited by compute (two speeds).
- No offload cliff. There is no separate RAM to spill into; a model over the limit simply does not fit well (labels).
- Quiet and efficient. Macs draw far less power than a high-end graphics card for the same model.
See every Mac we cover under Hardware, or press Detect on My hardware for yours.