Can an Apple M4 Pro 48GB run Llama 3.3 70B Instruct?

In practice

On paper the Apple M4 Pro 48GB has 48 GB, and Llama 3.3 70B Instruct at Q4_K_M needs 42.68 GB. macOS does not let the GPU use all of it: by default about 75% is available, which is 36 GB. So the model is just over the default limit and well inside the physical memory.

The limit is a setting, and one command raises it for the current session. It resets at reboot, and the memory it takes comes from the rest of the system, so keep other large programs closed while the model is loaded. Once it is raised the estimate is about 4.5 tokens per second.

The numbers

Llama 3.3 70B Instruct has 70.55 billion parameters across 80 layers. The Apple M4 Pro 48GB has 48 GB of unified memory at 273 GB/s, of which about 36 GB is left for a model once the operating system has taken its share.

4 of the 9 levels fit entirely. Levels marked degraded are listed for completeness, not as advice.

The settings we would use

sudo sysctl iogpu.wired_limit_mb=45056

macOS lets the GPU wire down about 75% of the memory by default, which is less than this model needs. The command raises the limit for this session and resets at reboot.

Once the limit is raised, Ollama and LM Studio use the extra memory without further settings. Run ollama ps and check that the processor column reads 100% GPU, and see the Mac and PC comparison for what unified memory does well and where it is slow.

If you would rather not change the limit

A smaller model from the same family is the straightforward answer. These fit on an Apple M4 Pro 48GB at a quantization worth using:

• Llama 3.1 8B Instruct, 8.03B at Q8_0, 9.32 GB, about 22.0 tokens/s
• Llama 3.2 3B Instruct, 3.21B at Q8_0, 4.39 GB, about 45.8 tokens/s
• Llama 3.2 1B Instruct, 1.24B at Q8_0, 1.79 GB, about 126 tokens/s

Models of this size are what a second GPU or a large-memory machine is for. The multi-GPU guide shows what 2 cards add, and the Mac and PC comparison shows where unified memory holds a model that no single card can.

What else an Apple M4 Pro 48GB runs well

For comparison, these are the other popular models that fit this device at a sound quantization, best pairings first.

The same model on nearby hardware

Memory decides whether Llama 3.3 70B Instruct loads and bandwidth decides how quickly it answers. Here is what a step down, a card with the same memory and a step up do with it.

Wait for the first word26 s

Then writes at4.5 tok/s

Whole answer1 min 4 s

YouWhy does my model use more memory when the conversation gets longer?

Model

Simulated from our estimate at a 512-token question, not a recording. Assumes nothing else is competing for the GPU.

Questions people ask

Keep going

Everything the Apple M4 Pro 48GB runs, the full breakdown for Llama 3.3 70B Instruct, how the numbers are made in the methodology, or a different pairing in the calculator.