How fast is Llama 3.1 405B on an Mac?

Around 3 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 405B need?

231.0 GiB for the weights at Q4_K_M, plus 3.9 GiB for an 8K-token KV cache and about 12.4 GiB of runtime overhead — 247.3 GiB in total.

Home › Llama 3.1 405B › M3 Ultra 512GB

Can an M3 Ultra 512GB Mac run Llama 3.1 405B?

M3 Ultra · 512 GB unified memory · 819 GB/s · Mac Studio (M3 Ultra)

Yes — Llama 3.1 405B runs on an M3 Ultra 512GB Mac.

At Q4_K_M it needs about 247.3 GiB, leaving 248.7 GiB spare out of the ~496.0 GiB macOS will let you use. Expect roughly 3 tokens/sec — too slow for interactive use.

Weights

231.0 GiB

405B params @ Q4_K_M

KV cache

3.9 GiB

8K context, FP16

Overhead

12.4 GiB

runtime + activations

Total needed

247.3 GiB

of ~496.0 GiB usable

Est. speed

3 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	754.4 GiB	796.8 GiB	1	No
Q8_0	400.8 GiB	425.5 GiB	2	Yes, but tight
Q6_K	311.2 GiB	331.5 GiB	2	Yes, comfortably
Q5_K_M	268.7 GiB	286.9 GiB	2	Yes, comfortably
Q4_K_M	231.0 GiB	247.3 GiB	3	Yes, comfortably
Q3_K_M	183.9 GiB	197.8 GiB	3	Yes, comfortably
Q2_K	141.4 GiB	153.3 GiB	4	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	1.0 GiB	244.1 GiB	Yes, comfortably	Yes, comfortably
4K	2.0 GiB	245.2 GiB	Yes, comfortably	Yes, comfortably
8K	3.9 GiB	247.3 GiB	Yes, comfortably	Yes, comfortably
16K	7.9 GiB	251.6 GiB	Yes, comfortably	Yes, comfortably
32K	15.8 GiB	260.0 GiB	Yes, comfortably	Yes, comfortably
128K	63.0 GiB	310.9 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 llama3.1:405b

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-3-1-405B-4bit --prompt "Hello"

Macs that run Llama 3.1 405B comfortably

The smallest configuration that handles it is the M3 Ultra 512GB (Mac Studio (M3 Ultra)).

Mac	Memory	Bandwidth	Tok/s
M3 Ultra 512GB	512 GB	819 GB/s	3

Common questions

Can an M3 Ultra 512GB Mac run Llama 3.1 405B?: Yes. At Q4_K_M it needs about 247.3 GiB of the roughly 496.0 GiB available, generating around 3 tokens per second.
How fast is Llama 3.1 405B on an Mac?: Around 3 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 405B need?: 231.0 GiB for the weights at Q4_K_M, plus 3.9 GiB for an 8K-token KV cache and about 12.4 GiB of runtime overhead — 247.3 GiB in total.