How fast is Gemma 2 27B on an Mac?

Around 6 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 120 GB/s.

How much memory does Gemma 2 27B need?

15.5 GiB for the weights at Q4_K_M, plus 2.9 GiB for an 8K-token KV cache and about 1.6 GiB of runtime overhead — 20.0 GiB in total.

Home › Gemma 2 27B › M4 32GB

Can an M4 32GB Mac run Gemma 2 27B?

M4 · 32 GB unified memory · 120 GB/s · MacBook Air 15" (M4), Mac mini (M4)

Yes — Gemma 2 27B runs on an M4 32GB Mac.

At Q4_K_M it needs about 20.0 GiB, leaving 7.2 GiB spare out of the ~27.2 GiB macOS will let you use. Expect roughly 6 tokens/sec — too slow for interactive use.

Weights

15.5 GiB

27.2B params @ Q4_K_M

KV cache

2.9 GiB

8K context, FP16

Overhead

1.6 GiB

runtime + activations

Total needed

20.0 GiB

of ~27.2 GiB usable

Est. speed

6 tok/s

120 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	50.7 GiB	56.9 GiB	2	No
Q8_0	26.9 GiB	31.9 GiB	3	No
Q6_K	20.9 GiB	25.6 GiB	4	Yes, but tight
Q5_K_M	18.0 GiB	22.6 GiB	5	Yes, comfortably
Q4_K_M	15.5 GiB	20.0 GiB	6	Yes, comfortably
Q3_K_M	12.3 GiB	16.6 GiB	7	Yes, comfortably
Q2_K	9.5 GiB	13.6 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	0.7 GiB	17.6 GiB	Yes, comfortably	Yes, comfortably
4K	1.4 GiB	18.4 GiB	Yes, comfortably	Yes, comfortably
8K	2.9 GiB	20.0 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 gemma2:27b

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-2-27B-4bit --prompt "Hello"

Macs that run Gemma 2 27B comfortably

The smallest configuration that handles it is the M1 Max 32GB (MacBook Pro 14" (M1 Max)).

Mac	Memory	Bandwidth	Tok/s
M1 Max 32GB	32 GB	400 GB/s	19
M2 Max 32GB	32 GB	400 GB/s	19
M4 32GB	32 GB	120 GB/s	6
M3 Max 36GB	36 GB	300 GB/s	14
M4 Max 36GB	36 GB	410 GB/s	20
M4 Pro 48GB	48 GB	273 GB/s	13
M4 Max 48GB	48 GB	546 GB/s	26
M1 Max 64GB	64 GB	400 GB/s	19
M2 Ultra 64GB	64 GB	800 GB/s	38
M4 Pro 64GB	64 GB	273 GB/s	13
M4 Max 64GB	64 GB	546 GB/s	26
M2 Max 96GB	96 GB	400 GB/s	19

Common questions

Can an M4 32GB Mac run Gemma 2 27B?: Yes. At Q4_K_M it needs about 20.0 GiB of the roughly 27.2 GiB available, generating around 6 tokens per second.
How fast is Gemma 2 27B on an Mac?: Around 6 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 120 GB/s.
How much memory does Gemma 2 27B need?: 15.5 GiB for the weights at Q4_K_M, plus 2.9 GiB for an 8K-token KV cache and about 1.6 GiB of runtime overhead — 20.0 GiB in total.