How fast is Llama 3.1 8B on an Mac?

Around 133 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s.

How much memory does Llama 3.1 8B need?

4.6 GiB for the weights at Q4_K_M, plus 1.0 GiB for an 8K-token KV cache and about 1.0 GiB of runtime overhead — 6.6 GiB in total.

Home › Llama 3.1 8B › M3 Ultra 96GB

Can an M3 Ultra 96GB Mac run Llama 3.1 8B?

M3 Ultra · 96 GB unified memory · 819 GB/s · Mac Studio (M3 Ultra)

Yes — Llama 3.1 8B runs on an M3 Ultra 96GB Mac.

At Q4_K_M it needs about 6.6 GiB, leaving 75.0 GiB spare out of the ~81.6 GiB macOS will let you use. Expect roughly 133 tokens/sec — faster than you can read.

Weights

4.6 GiB

8.03B params @ Q4_K_M

KV cache

1.0 GiB

8K context, FP16

Overhead

1.0 GiB

runtime + activations

Total needed

6.6 GiB

of ~81.6 GiB usable

Est. speed

133 tok/s

819 GB/s bandwidth

Max context

128K 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	15.0 GiB	17.5 GiB	41	Yes, comfortably
Q8_0	7.9 GiB	10.1 GiB	77	Yes, comfortably
Q6_K	6.2 GiB	8.3 GiB	99	Yes, comfortably
Q5_K_M	5.3 GiB	7.4 GiB	115	Yes, comfortably
Q4_K_M	4.6 GiB	6.6 GiB	133	Yes, comfortably
Q3_K_M	3.6 GiB	5.6 GiB	167	Yes, comfortably
Q2_K	2.8 GiB	4.7 GiB	218	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	0.3 GiB	5.6 GiB	Yes, comfortably	Yes, comfortably
4K	0.5 GiB	6.0 GiB	Yes, comfortably	Yes, comfortably
8K	1.0 GiB	6.6 GiB	Yes, comfortably	Yes, comfortably
16K	2.0 GiB	7.9 GiB	Yes, comfortably	Yes, comfortably
32K	4.0 GiB	10.5 GiB	Yes, comfortably	Yes, comfortably
128K	16.0 GiB	26.1 GiB	Yes, comfortably	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 llama3.1:8b

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-3-1-8B-4bit --prompt "Hello"

Macs that run Llama 3.1 8B comfortably

The smallest configuration that handles it is the M1 16GB (MacBook Air (M1)).

Mac	Memory	Bandwidth	Tok/s
M1 16GB	16 GB	68 GB/s	11
M1 Pro 16GB	16 GB	200 GB/s	33
M2 16GB	16 GB	100 GB/s	16
M4 16GB	16 GB	120 GB/s	20
M3 Pro 18GB	18 GB	150 GB/s	24
M2 24GB	24 GB	100 GB/s	16
M4 24GB	24 GB	120 GB/s	20
M4 Pro 24GB	24 GB	273 GB/s	44
M1 Max 32GB	32 GB	400 GB/s	65
M2 Max 32GB	32 GB	400 GB/s	65
M4 32GB	32 GB	120 GB/s	20
M3 Max 36GB	36 GB	300 GB/s	49

Common questions

Can an M3 Ultra 96GB Mac run Llama 3.1 8B?: Yes. At Q4_K_M it needs about 6.6 GiB of the roughly 81.6 GiB available, generating around 133 tokens per second.
How fast is Llama 3.1 8B on an Mac?: Around 133 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s.
How much memory does Llama 3.1 8B need?: 4.6 GiB for the weights at Q4_K_M, plus 1.0 GiB for an 8K-token KV cache and about 1.0 GiB of runtime overhead — 6.6 GiB in total.