How fast is Llama 3.3 70B on an Mac?

Around 7 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 400 GB/s.

How much memory does Llama 3.3 70B need?

40.3 GiB for the weights at Q4_K_M, plus 2.5 GiB for an 8K-token KV cache and about 2.8 GiB of runtime overhead — 45.6 GiB in total.

Home › Llama 3.3 70B › M2 Max 96GB

Can an M2 Max 96GB Mac run Llama 3.3 70B?

M2 Max · 96 GB unified memory · 400 GB/s · MacBook Pro 16" (M2 Max), Mac Studio (M2 Max)

Yes — Llama 3.3 70B runs on an M2 Max 96GB Mac.

At Q4_K_M it needs about 45.6 GiB, leaving 36.0 GiB spare out of the ~81.6 GiB macOS will let you use. Expect roughly 7 tokens/sec — slow but workable for background tasks.

Weights

40.3 GiB

70.6B params @ Q4_K_M

KV cache

2.5 GiB

8K context, FP16

Overhead

2.8 GiB

runtime + activations

Total needed

45.6 GiB

of ~81.6 GiB usable

Est. speed

7 tok/s

400 GB/s bandwidth

Max context

64K tokens

at Q4_K_M

Where the memory goes

Three things occupy memory when a model is loaded: the weights themselves, the KV cache that holds the conversation, and the runtime's own working buffers. Only the first is fixed — the KV cache grows with every token in your context.

Quantisation	Weights	Total	Tok/s	Fits?
FP16	131.5 GiB	141.4 GiB	2	No
Q8_0	69.9 GiB	76.7 GiB	4	Yes, but tight
Q6_K	54.2 GiB	60.3 GiB	5	Yes, comfortably
Q5_K_M	46.8 GiB	52.5 GiB	6	Yes, comfortably
Q4_K_M	40.3 GiB	45.6 GiB	7	Yes, comfortably
Q3_K_M	32.1 GiB	37.0 GiB	9	Yes, comfortably
Q2_K	24.7 GiB	29.2 GiB	12	Yes, comfortably

The default. Best quality-per-gigabyte for most people.

How long a context fits

The KV cache is often what breaks a setup that looked fine at load time. Quantising it to 8-bit roughly halves its footprint with little measurable quality cost.

Context	KV cache	Total	FP16 KV	Q8 KV
2K	0.6 GiB	43.5 GiB	Yes, comfortably	Yes, comfortably
4K	1.3 GiB	44.2 GiB	Yes, comfortably	Yes, comfortably
8K	2.5 GiB	45.6 GiB	Yes, comfortably	Yes, comfortably
16K	5.0 GiB	48.4 GiB	Yes, comfortably	Yes, comfortably
32K	10.0 GiB	54.0 GiB	Yes, comfortably	Yes, comfortably
128K	40.0 GiB	87.6 GiB	No	Yes, comfortably

How to run it

Ollama is the shortest path. It picks a quantisation automatically — usually Q4_K_M, which is what the numbers above assume.

brew install ollama
ollama serve &
ollama run llama3.3:70b

For MLX — Apple's own array framework, typically a little faster than llama.cpp on Apple Silicon:

pip install mlx-lm
mlx_lm.generate --model mlx-community/Llama-3-3-70B-4bit --prompt "Hello"

Macs that run Llama 3.3 70B comfortably

The smallest configuration that handles it is the M1 Max 64GB (MacBook Pro 16" (M1 Max)).

Mac	Memory	Bandwidth	Tok/s
M1 Max 64GB	64 GB	400 GB/s	7
M2 Ultra 64GB	64 GB	800 GB/s	15
M4 Pro 64GB	64 GB	273 GB/s	5
M4 Max 64GB	64 GB	546 GB/s	10
M2 Max 96GB	96 GB	400 GB/s	7
M3 Ultra 96GB	96 GB	819 GB/s	15
M3 Max 128GB	128 GB	400 GB/s	7
M4 Max 128GB	128 GB	546 GB/s	10
M2 Ultra 192GB	192 GB	800 GB/s	15
M3 Ultra 256GB	256 GB	819 GB/s	15
M3 Ultra 512GB	512 GB	819 GB/s	15

Common questions

Can an M2 Max 96GB Mac run Llama 3.3 70B?: Yes. At Q4_K_M it needs about 45.6 GiB of the roughly 81.6 GiB available, generating around 7 tokens per second.
How fast is Llama 3.3 70B on an Mac?: Around 7 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 400 GB/s.
How much memory does Llama 3.3 70B need?: 40.3 GiB for the weights at Q4_K_M, plus 2.5 GiB for an 8K-token KV cache and about 2.8 GiB of runtime overhead — 45.6 GiB in total.