How fast is Gemma 3 4B on an Mac?

Around 21 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 68 GB/s.

How much memory does Gemma 3 4B need?

2.5 GiB for the weights at Q4_K_M, plus 1.1 GiB for an 8K-token KV cache and about 0.9 GiB of runtime overhead — 4.4 GiB in total.

Home › Gemma 3 4B › M1 8GB

Can an M1 8GB Mac run Gemma 3 4B?

M1 · 8 GB unified memory · 68 GB/s · MacBook Air (M1), MacBook Pro 13" (M1)

Just barely — and you will feel it.

Gemma 3 4B at Q4_K_M needs about 4.4 GiB against roughly 5.0 GiB usable — that is 89% of your budget. It will load, but close your browser first, keep the context short, and expect memory pressure. Around 21 tokens/sec.

Weights

2.5 GiB

4.3B params @ Q4_K_M

KV cache

1.1 GiB

8K context, FP16

Overhead

0.9 GiB

runtime + activations

Total needed

4.4 GiB

of ~5.0 GiB usable

Est. speed

21 tok/s

68 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	8.0 GiB	10.3 GiB	6	No
Q8_0	4.3 GiB	6.3 GiB	12	No
Q6_K	3.3 GiB	5.3 GiB	15	No
Q5_K_M	2.9 GiB	4.9 GiB	18	Yes, but tight
Q4_K_M	2.5 GiB	4.4 GiB	21	Yes, but tight
Q3_K_M	2.0 GiB	3.9 GiB	26	Yes, comfortably
Q2_K	1.5 GiB	3.4 GiB	34	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	3.4 GiB	Yes, comfortably	Yes, comfortably
4K	0.5 GiB	3.8 GiB	Yes, comfortably	Yes, comfortably
8K	1.1 GiB	4.4 GiB	Yes, but tight	Yes, comfortably
16K	2.1 GiB	5.8 GiB	No	Yes, but tight
32K	4.3 GiB	8.5 GiB	No	No
128K	17.0 GiB	24.9 GiB	No	No

What to run instead on an M1 8GB Mac

These are the largest models that fit comfortably on this machine at Q4_K_M with an 8K context.

Model	Size	Total	Tok/s
Llama 3.2 3B	3.21B	3.6 GiB	28	Yes, comfortably
Qwen2.5 3B	3.09B	2.9 GiB	29	Yes, comfortably
Gemma 2 2B	2.61B	3.0 GiB	34	Yes, comfortably
Qwen2.5 1.5B	1.54B	1.9 GiB	58	Yes, comfortably
Llama 3.2 1B	1.24B	1.8 GiB	72	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 gemma3:4b

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-3-4B-4bit --prompt "Hello"

You are close to the limitIf it stutters or gets killed, raise the Metal memory cap before loading: sudo sysctl iogpu.wired_limit_mb=6963. This resets on reboot. Quantising the KV cache (--kv-cache-type q8_0 in llama.cpp) buys back 0.5 GiB.

Macs that run Gemma 3 4B 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	21
M1 Pro 16GB	16 GB	200 GB/s	61
M2 16GB	16 GB	100 GB/s	30
M4 16GB	16 GB	120 GB/s	36
M3 Pro 18GB	18 GB	150 GB/s	46
M2 24GB	24 GB	100 GB/s	30
M4 24GB	24 GB	120 GB/s	36
M4 Pro 24GB	24 GB	273 GB/s	83
M1 Max 32GB	32 GB	400 GB/s	121
M2 Max 32GB	32 GB	400 GB/s	121
M4 32GB	32 GB	120 GB/s	36
M3 Max 36GB	36 GB	300 GB/s	91

Common questions

Can an M1 8GB Mac run Gemma 3 4B?: Yes. At Q4_K_M it needs about 4.4 GiB of the roughly 5.0 GiB available, generating around 21 tokens per second.
How fast is Gemma 3 4B on an Mac?: Around 21 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 68 GB/s.
How much memory does Gemma 3 4B need?: 2.5 GiB for the weights at Q4_K_M, plus 1.1 GiB for an 8K-token KV cache and about 0.9 GiB of runtime overhead — 4.4 GiB in total.