How fast is Llama 3.1 405B on an Mac?

Around 2 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 546 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 › M4 Max 128GB

Can an M4 Max 128GB Mac run Llama 3.1 405B?

M4 Max · 128 GB unified memory · 546 GB/s · MacBook Pro 16" (M4 Max), Mac Studio (M4 Max)

No — not at Q4_K_M.

Llama 3.1 405B needs about 247.3 GiB but an M4 Max 128GB Mac has only ~112.0 GiB available for a model — you are 135.3 GiB short. macOS will swap to SSD and generation will crawl.

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 ~112.0 GiB usable

Est. speed

2 tok/s

546 GB/s bandwidth

Max context

n/a

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	1	No
Q6_K	311.2 GiB	331.5 GiB	1	No
Q5_K_M	268.7 GiB	286.9 GiB	2	No
Q4_K_M	231.0 GiB	247.3 GiB	2	No
Q3_K_M	183.9 GiB	197.8 GiB	2	No
Q2_K	141.4 GiB	153.3 GiB	3	No

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	No	No
4K	2.0 GiB	245.2 GiB	No	No
8K	3.9 GiB	247.3 GiB	No	No
16K	7.9 GiB	251.6 GiB	No	No
32K	15.8 GiB	260.0 GiB	No	No
128K	63.0 GiB	310.9 GiB	No	No

What to run instead on an M4 Max 128GB Mac

These are the largest models that fit comfortably on this machine at Q4_K_M with an 8K context.

Model	Size	Total	Tok/s
gpt-oss-120b	116.8B	71.3 GiB	87	Yes, comfortably
Qwen2.5 72B	72.7B	46.8 GiB	10	Yes, comfortably
DeepSeek-R1-Distill-Llama-70B	70.6B	45.6 GiB	10	Yes, comfortably
Llama 3.3 70B	70.6B	45.6 GiB	10	Yes, comfortably
Mixtral 8x7B	46.7B	29.8 GiB	35	Yes, comfortably
DeepSeek-R1-Distill-Qwen-32B	32.8B	22.4 GiB	22	Yes, comfortably

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 M4 Max 128GB Mac run Llama 3.1 405B?: No. At Q4_K_M it needs about 247.3 GiB, and an M4 Max 128GB Mac has roughly 112.0 GiB available for a model after macOS takes its share.
How fast is Llama 3.1 405B on an Mac?: Around 2 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 546 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.