How fast is Gemma 2 2B on an Mac?

Around 60 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 2B need?

1.5 GiB for the weights at Q4_K_M, plus 0.7 GiB for an 8K-token KV cache and about 0.9 GiB of runtime overhead — 3.0 GiB in total.

Home › Gemma 2 2B › M4 32GB

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

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

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

At Q4_K_M it needs about 3.0 GiB, leaving 24.2 GiB spare out of the ~27.2 GiB macOS will let you use. Expect roughly 60 tokens/sec — faster than you can read.

Weights

1.5 GiB

2.61B params @ Q4_K_M

KV cache

0.7 GiB

8K context, FP16

Overhead

0.9 GiB

runtime + activations

Total needed

3.0 GiB

of ~27.2 GiB usable

Est. speed

60 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	4.9 GiB	6.6 GiB	18	Yes, comfortably
Q8_0	2.6 GiB	4.2 GiB	35	Yes, comfortably
Q6_K	2.0 GiB	3.6 GiB	45	Yes, comfortably
Q5_K_M	1.7 GiB	3.3 GiB	52	Yes, comfortably
Q4_K_M	1.5 GiB	3.0 GiB	60	Yes, comfortably
Q3_K_M	1.2 GiB	2.7 GiB	75	Yes, comfortably
Q2_K	0.9 GiB	2.4 GiB	98	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	0.2 GiB	2.3 GiB	Yes, comfortably	Yes, comfortably
4K	0.3 GiB	2.5 GiB	Yes, comfortably	Yes, comfortably
8K	0.7 GiB	3.0 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 gemma2:2b

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

Macs that run Gemma 2 2B comfortably

The smallest configuration that handles it is the M1 8GB (MacBook Air (M1)).

Mac	Memory	Bandwidth	Tok/s
M1 8GB	8 GB	68 GB/s	34
M1 16GB	16 GB	68 GB/s	34
M1 Pro 16GB	16 GB	200 GB/s	100
M2 16GB	16 GB	100 GB/s	50
M4 16GB	16 GB	120 GB/s	60
M3 Pro 18GB	18 GB	150 GB/s	75
M2 24GB	24 GB	100 GB/s	50
M4 24GB	24 GB	120 GB/s	60
M4 Pro 24GB	24 GB	273 GB/s	137
M1 Max 32GB	32 GB	400 GB/s	200
M2 Max 32GB	32 GB	400 GB/s	200
M4 32GB	32 GB	120 GB/s	60

Common questions

Can an M4 32GB Mac run Gemma 2 2B?: Yes. At Q4_K_M it needs about 3.0 GiB of the roughly 27.2 GiB available, generating around 60 tokens per second.
How fast is Gemma 2 2B on an Mac?: Around 60 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 2B need?: 1.5 GiB for the weights at Q4_K_M, plus 0.7 GiB for an 8K-token KV cache and about 0.9 GiB of runtime overhead — 3.0 GiB in total.