How fast is Gemma 2 9B on an Mac?

Around 42 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 300 GB/s.

How much memory does Gemma 2 9B need?

5.3 GiB for the weights at Q4_K_M, plus 2.6 GiB for an 8K-token KV cache and about 1.1 GiB of runtime overhead — 9.0 GiB in total.

Home › Gemma 2 9B › M3 Max 36GB

Can an M3 Max 36GB Mac run Gemma 2 9B?

M3 Max · 36 GB unified memory · 300 GB/s · MacBook Pro 14" (M3 Max), MacBook Pro 16" (M3 Max)

Yes — Gemma 2 9B runs on an M3 Max 36GB Mac.

At Q4_K_M it needs about 9.0 GiB, leaving 21.6 GiB spare out of the ~30.6 GiB macOS will let you use. Expect roughly 42 tokens/sec — faster than you can read.

Weights

5.3 GiB

9.24B params @ Q4_K_M

KV cache

2.6 GiB

8K context, FP16

Overhead

1.1 GiB

runtime + activations

Total needed

9.0 GiB

of ~30.6 GiB usable

Est. speed

42 tok/s

300 GB/s bandwidth

Max context

8K 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	17.2 GiB	21.5 GiB	13	Yes, comfortably
Q8_0	9.1 GiB	13.0 GiB	24	Yes, comfortably
Q6_K	7.1 GiB	10.9 GiB	31	Yes, comfortably
Q5_K_M	6.1 GiB	9.9 GiB	36	Yes, comfortably
Q4_K_M	5.3 GiB	9.0 GiB	42	Yes, comfortably
Q3_K_M	4.2 GiB	7.8 GiB	53	Yes, comfortably
Q2_K	3.2 GiB	6.8 GiB	69	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	0.7 GiB	6.8 GiB	Yes, comfortably	Yes, comfortably
4K	1.3 GiB	7.5 GiB	Yes, comfortably	Yes, comfortably
8K	2.6 GiB	9.0 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 gemma2:9b

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/Gemma-2-9B-4bit --prompt "Hello"

Macs that run Gemma 2 9B 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	10
M1 Pro 16GB	16 GB	200 GB/s	28
M2 16GB	16 GB	100 GB/s	14
M4 16GB	16 GB	120 GB/s	17
M3 Pro 18GB	18 GB	150 GB/s	21
M2 24GB	24 GB	100 GB/s	14
M4 24GB	24 GB	120 GB/s	17
M4 Pro 24GB	24 GB	273 GB/s	39
M1 Max 32GB	32 GB	400 GB/s	57
M2 Max 32GB	32 GB	400 GB/s	57
M4 32GB	32 GB	120 GB/s	17
M3 Max 36GB	36 GB	300 GB/s	42

Common questions

Can an M3 Max 36GB Mac run Gemma 2 9B?: Yes. At Q4_K_M it needs about 9.0 GiB of the roughly 30.6 GiB available, generating around 42 tokens per second.
How fast is Gemma 2 9B on an Mac?: Around 42 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 300 GB/s.
How much memory does Gemma 2 9B need?: 5.3 GiB for the weights at Q4_K_M, plus 2.6 GiB for an 8K-token KV cache and about 1.1 GiB of runtime overhead — 9.0 GiB in total.