How fast is Mistral Small 3 24B on an Mac?

Around 6 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 100 GB/s.

How much memory does Mistral Small 3 24B need?

13.5 GiB for the weights at Q4_K_M, plus 1.3 GiB for an 8K-token KV cache and about 1.5 GiB of runtime overhead — 16.2 GiB in total.

Home › Mistral Small 3 24B › M2 24GB

Can an M2 24GB Mac run Mistral Small 3 24B?

M2 · 24 GB unified memory · 100 GB/s · MacBook Air 15" (M2), Mac mini (M2)

Yes — Mistral Small 3 24B runs on an M2 24GB Mac.

At Q4_K_M it needs about 16.2 GiB, leaving 4.2 GiB spare out of the ~20.4 GiB macOS will let you use. Expect roughly 6 tokens/sec — too slow for interactive use.

Weights

13.5 GiB

23.6B params @ Q4_K_M

KV cache

1.3 GiB

8K context, FP16

Overhead

1.5 GiB

runtime + activations

Total needed

16.2 GiB

of ~20.4 GiB usable

Est. speed

6 tok/s

100 GB/s bandwidth

Max context

16K 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	44.0 GiB	48.2 GiB	2	No
Q8_0	23.4 GiB	26.6 GiB	3	No
Q6_K	18.1 GiB	21.1 GiB	4	No
Q5_K_M	15.7 GiB	18.5 GiB	5	Yes, but tight
Q4_K_M	13.5 GiB	16.2 GiB	6	Yes, comfortably
Q3_K_M	10.7 GiB	13.3 GiB	7	Yes, comfortably
Q2_K	8.2 GiB	10.7 GiB	9	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	15.0 GiB	Yes, comfortably	Yes, comfortably
4K	0.6 GiB	15.4 GiB	Yes, comfortably	Yes, comfortably
8K	1.3 GiB	16.2 GiB	Yes, comfortably	Yes, comfortably
16K	2.5 GiB	17.7 GiB	Yes, but tight	Yes, comfortably
32K	5.0 GiB	20.8 GiB	No	Yes, but tight

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 mistral-small

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/Mistral-Small-3-24B-4bit --prompt "Hello"

Macs that run Mistral Small 3 24B comfortably

The smallest configuration that handles it is the M2 24GB (MacBook Air 15" (M2)).

Mac	Memory	Bandwidth	Tok/s
M2 24GB	24 GB	100 GB/s	6
M4 24GB	24 GB	120 GB/s	7
M4 Pro 24GB	24 GB	273 GB/s	15
M1 Max 32GB	32 GB	400 GB/s	22
M2 Max 32GB	32 GB	400 GB/s	22
M4 32GB	32 GB	120 GB/s	7
M3 Max 36GB	36 GB	300 GB/s	17
M4 Max 36GB	36 GB	410 GB/s	23
M4 Pro 48GB	48 GB	273 GB/s	15
M4 Max 48GB	48 GB	546 GB/s	30
M1 Max 64GB	64 GB	400 GB/s	22
M2 Ultra 64GB	64 GB	800 GB/s	44

Common questions

Can an M2 24GB Mac run Mistral Small 3 24B?: Yes. At Q4_K_M it needs about 16.2 GiB of the roughly 20.4 GiB available, generating around 6 tokens per second.
How fast is Mistral Small 3 24B on an Mac?: Around 6 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 100 GB/s.
How much memory does Mistral Small 3 24B need?: 13.5 GiB for the weights at Q4_K_M, plus 1.3 GiB for an 8K-token KV cache and about 1.5 GiB of runtime overhead — 16.2 GiB in total.