The best Mac for running local AI models
Running a large language model on your own Mac keeps your prompts at home and costs nothing per question. Tools such as Ollama, LM Studio and llama.cpp make it a few clicks, but the Mac you pick decides which models you can run and how fast they answer.
Two numbers matter more than anything else. Unified memory decides which models fit: the whole model has to stay in memory while it runs, and on Apple silicon the GPU shares that memory with the rest of the Mac. Memory bandwidth decides speed: for every word the model writes, the Mac reads the whole model once, so a Mac that moves more gigabytes per second answers faster.
This ranking puts the Mac that runs the largest model at a usable speed first and breaks ties by how quickly it answers. Each card lists what every memory option can run, and the speeds are estimates from Apple's published memory bandwidth, not measurements.
Prices are Apple's price for each configuration, not used prices.
Ranked for this use
How we judge
Which models fit. We assume 4-bit quantised models, the usual choice for running at home, which take about 0.6 GB of memory per billion parameters. We keep 6 GB free for macOS and your other apps. A model fits if the rest of the memory is enough for it.
How fast it answers (estimate). Writing each token means reading the whole model from memory, so we divide Apple's published memory bandwidth by the size of the model and assume 60% of that peak is reached in practice. Real speed depends on the app, the model and the length of the conversation. Where a published measurement exists for a chip and model, it replaces the estimate.
Usable means at least 10 tokens per second, roughly as fast as you read. We check these reference models: Llama 3.1 8B, Qwen 2.5 14B, Qwen 2.5 32B, Llama 3.3 70B, Mistral Large 123B, and Llama 3.1 405B.
Great: runs a 70B model or larger at a usable speed. Good enough: runs at least a 14B model at a usable speed. Not ideal: only smaller models. Unknown: Apple publishes no memory bandwidth for that Mac, so we don't guess.
Memory and bandwidth come from Apple's technical specifications for each Mac. Only the configurations listed here are ranked; most Macs could be ordered with more memory.
Questions and answers
How much memory do I need to run Llama 70B on a Mac?
By our sizing, a 4-bit 70B model takes about 42 GB, and with room for macOS you need around 48 GB in total. In practice that means a Mac with 64 GB of unified memory or more. Memory alone is not enough, though: to answer at a usable speed, it also needs the memory bandwidth of an Ultra chip.
Is 16 GB enough to run AI models on a Mac?
For smaller models, yes. An 8B model such as Llama 3.1 8B needs about 10.8 GB with macOS included, so it fits in 16 GB and is useful for chat, summaries and simple coding help. A 14B model still just fits. A 32B model needs about 25.2 GB, so 32 GB, and the largest open models need 64 GB or more.
Why does memory bandwidth matter more than the chip for local AI?
Because generating text is mostly waiting for memory. For every token, the Mac reads the whole model once, so tokens per second is roughly bandwidth divided by model size. A Max or Ultra chip with several hundred GB/s answers far faster than a base chip with the same amount of memory. Extra GPU cores help most while the model reads your prompt.
Mac mini or Mac Studio for local AI?
A Mac mini with a Pro chip and enough memory runs 8B and 14B models comfortably. For 32B models and up, the Max and Ultra chips in the Mac Studio have the memory bandwidth to keep them usable, and in our ranking only Ultra chips run 70B models at a usable speed.
Which apps run AI models on a Mac?
Ollama and LM Studio are the easiest start: download a model and chat. Under the hood they use engines such as llama.cpp and Apple's MLX, which run the model on the Mac's GPU. A model's file size is a good guide to how much memory it needs.
Are the speeds on this page measured?
Not yet. They are estimates from Apple's published memory bandwidth, assuming the Mac reaches 60% of its peak, and they are labelled as estimates. Long conversations and large prompts run slower than the estimate. When a published measurement exists for a chip and model, we show it instead.