Can an Apple M3 Max 36GB run Llama 3.3 70B Instruct?

Q: Can an Apple M3 Max 36GB run Llama 3.3 70B Instruct?

Yes, with a setting. It needs 28.14 GB at IQ3_XXS and macOS allows 25.2 GB by default. Raise the limit and it runs at about 10.0 tokens per second.

In practice

It loads, and we would not call that running it. The only quantization that fits the Apple M3 Max 36GB is IQ3_XXS, which compresses the weights far enough that Llama 3.3 70B Instruct loses a visible amount of its ability, so we do not recommend it. At Q4_K_M it needs 42.68 GB against the 25.2 GB available.

A smaller model at a proper quantization is a better answer than a larger one squeezed in, and the next section lists the ones that fit.

The numbers

Llama 3.3 70B Instruct has 70.55 billion parameters across 80 layers. The Apple M3 Max 36GB has 36 GB of unified memory at 400 GB/s, of which about 25.2 GB is left for a model once the operating system has taken its share.

Quantization	Total needed	Fits	Speed
Q8_0	72.82 GB	no	n/a
Q6_K	56.89 GB	no	n/a
Q5_K_M	49.74 GB	no	n/a
Q4_K_M	42.68 GB	no	n/a
IQ4_XS	37.92 GB	no	n/a
Q3_K_M	35.12 GB	no	n/a
IQ3_XXS degraded	28.14 GB	with raised limit	10.0 t/s
Q2_K degraded	30.53 GB	with raised limit	9.2 t/s
IQ2_XXS degraded	19.93 GB	yes	14.2 t/s

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

If you would rather not change the limit

A smaller model from the same family is the straightforward answer. These fit on an Apple M3 Max 36GB at a quantization worth using:

Llama 3.1 8B Instruct, 8.03B at Q8_0, 9.32 GB, about 32.3 tokens/s
Llama 3.2 3B Instruct, 3.21B at Q8_0, 4.39 GB, about 67.1 tokens/s
Llama 3.2 1B Instruct, 1.24B at Q8_0, 1.79 GB, about 184 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 M3 Max 36GB runs well

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

Model	Quantization	Memory	Speed
Gemma 4 12B Instruct	Q8_0	13.52 GB	20.7 t/s
GPT OSS 20B	MXFP4	11.81 GB	66.0 t/s
Qwen3 14B	Q8_0	16.28 GB	18.0 t/s
Qwen3.5 9B	Q8_0	10.22 GB	30.8 t/s

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.

Device	Memory	Bandwidth	Result
GeForce RTX 5090	32 GB	1792 GB/s	does not fit
Apple M4 Pro 48GB	48 GB	273 GB/s	does not fit

See how fast it feels

Apple M3 Max 36GB running Llama 3.3 70B Instruct at IQ3_XXS

Wait for the first word17 s

Then writes at10 tok/s

Whole answer34 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

How much VRAM does Llama 3.3 70B Instruct need on Apple M3 Max 36GB?

At Q4_K_M with an 8k context window it needs 42.68 GB in total: 39.67 GB of weights, 2.5 GB of KV cache and 0.51 GB of runtime overhead. The Apple M3 Max 36GB makes about 25.2 GB available once the operating system has taken its share.

Can I run Llama 3.3 70B Instruct on Apple M3 Max 36GB anyway?

Yes, after raising the macOS GPU memory limit, which one sysctl command does. Without it the model needs 28.14 GB and the default limit allows 25.2 GB.

Keep going

Everything the Apple M3 Max 36GB 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.