How fast is Mixtral 8x7B on an Mac?

Around 52 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s. Because this is a mixture-of-experts model only 12.9B of its 46.7B parameters are read per token, which is why it is faster than its size suggests.

How much memory does Mixtral 8x7B need?

26.6 GiB for the weights at Q4_K_M, plus 1.0 GiB for an 8K-token KV cache and about 2.1 GiB of runtime overhead — 29.8 GiB in total.

Home › Mixtral 8x7B › M3 Ultra 96GB

Can an M3 Ultra 96GB Mac run Mixtral 8x7B?

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

Yes — Mixtral 8x7B runs on an M3 Ultra 96GB Mac.

At Q4_K_M it needs about 29.8 GiB, leaving 51.8 GiB spare out of the ~81.6 GiB macOS will let you use. Expect roughly 52 tokens/sec — faster than you can read.

Weights

26.6 GiB

46.7B params @ Q4_K_M

KV cache

1.0 GiB

8K context, FP16

Overhead

2.1 GiB

runtime + activations

Total needed

29.8 GiB

of ~81.6 GiB usable

Est. speed

52 tok/s

819 GB/s bandwidth

Max context

32K 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	87.0 GiB	93.1 GiB	16	No
Q8_0	46.2 GiB	50.3 GiB	30	Yes, comfortably
Q6_K	35.9 GiB	39.5 GiB	38	Yes, comfortably
Q5_K_M	31.0 GiB	34.3 GiB	45	Yes, comfortably
Q4_K_M	26.6 GiB	29.8 GiB	52	Yes, comfortably
Q3_K_M	21.2 GiB	24.1 GiB	65	Yes, comfortably
Q2_K	16.3 GiB	18.9 GiB	85	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	28.8 GiB	Yes, comfortably	Yes, comfortably
4K	0.5 GiB	29.1 GiB	Yes, comfortably	Yes, comfortably
8K	1.0 GiB	29.8 GiB	Yes, comfortably	Yes, comfortably
16K	2.0 GiB	31.1 GiB	Yes, comfortably	Yes, comfortably
32K	4.0 GiB	33.7 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 mixtral:8x7b

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/Mixtral-8x7B-4bit --prompt "Hello"

Macs that run Mixtral 8x7B comfortably

The smallest configuration that handles it is the M4 Pro 48GB (Mac mini (M4 Pro)).

Mac	Memory	Bandwidth	Tok/s
M4 Pro 48GB	48 GB	273 GB/s	17
M4 Max 48GB	48 GB	546 GB/s	35
M1 Max 64GB	64 GB	400 GB/s	25
M2 Ultra 64GB	64 GB	800 GB/s	51
M4 Pro 64GB	64 GB	273 GB/s	17
M4 Max 64GB	64 GB	546 GB/s	35
M2 Max 96GB	96 GB	400 GB/s	25
M3 Ultra 96GB	96 GB	819 GB/s	52
M3 Max 128GB	128 GB	400 GB/s	25
M4 Max 128GB	128 GB	546 GB/s	35
M2 Ultra 192GB	192 GB	800 GB/s	51
M3 Ultra 256GB	256 GB	819 GB/s	52

Common questions

Can an M3 Ultra 96GB Mac run Mixtral 8x7B?: Yes. At Q4_K_M it needs about 29.8 GiB of the roughly 81.6 GiB available, generating around 52 tokens per second.
How fast is Mixtral 8x7B on an Mac?: Around 52 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s. Because this is a mixture-of-experts model only 12.9B of its 46.7B parameters are read per token, which is why it is faster than its size suggests.
How much memory does Mixtral 8x7B need?: 26.6 GiB for the weights at Q4_K_M, plus 1.0 GiB for an 8K-token KV cache and about 2.1 GiB of runtime overhead — 29.8 GiB in total.