How fast is Gemma 3 12B on an Mac?

Around 21 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 200 GB/s.

How much memory does Gemma 3 12B need?

7.0 GiB for the weights at Q4_K_M, plus 3.0 GiB for an 8K-token KV cache and about 1.1 GiB of runtime overhead — 11.1 GiB in total.

Home › Gemma 3 12B › M1 Pro 16GB

Can an M1 Pro 16GB Mac run Gemma 3 12B?

M1 Pro · 16 GB unified memory · 200 GB/s · MacBook Pro 14" (M1 Pro), MacBook Pro 16" (M1 Pro)

Just barely — and you will feel it.

Gemma 3 12B at Q4_K_M needs about 11.1 GiB against roughly 13.0 GiB usable — that is 85% of your budget. It will load, but close your browser first, keep the context short, and expect memory pressure. Around 21 tokens/sec.

Weights

7.0 GiB

12.2B params @ Q4_K_M

KV cache

3.0 GiB

8K context, FP16

Overhead

1.1 GiB

runtime + activations

Total needed

11.1 GiB

of ~13.0 GiB usable

Est. speed

21 tok/s

200 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	22.7 GiB	27.7 GiB	7	No
Q8_0	12.1 GiB	16.5 GiB	12	No
Q6_K	9.4 GiB	13.6 GiB	16	No
Q5_K_M	8.1 GiB	12.3 GiB	18	Yes, but tight
Q4_K_M	7.0 GiB	11.1 GiB	21	Yes, but tight
Q3_K_M	5.5 GiB	9.6 GiB	27	Yes, comfortably
Q2_K	4.3 GiB	8.3 GiB	35	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	0.8 GiB	8.6 GiB	Yes, comfortably	Yes, comfortably
4K	1.5 GiB	9.5 GiB	Yes, comfortably	Yes, comfortably
8K	3.0 GiB	11.1 GiB	Yes, but tight	Yes, comfortably
16K	6.0 GiB	14.4 GiB	No	Yes, but tight
32K	12.0 GiB	21.0 GiB	No	No
128K	48.0 GiB	60.6 GiB	No	No

What to run instead on an M1 Pro 16GB 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
Qwen3 14B	14.8B	10.9 GiB	18	Yes, comfortably
Gemma 2 9B	9.24B	9.0 GiB	28	Yes, comfortably
Qwen3 8B	8.2B	6.8 GiB	32	Yes, comfortably
Llama 3.1 8B	8.03B	6.6 GiB	33	Yes, comfortably
DeepSeek-R1-Distill-Qwen-7B	7.62B	5.8 GiB	34	Yes, comfortably
Qwen2.5 7B	7.62B	5.8 GiB	34	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 gemma3:12b

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-3-12B-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=13926. This resets on reboot. Quantising the KV cache (--kv-cache-type q8_0 in llama.cpp) buys back 1.4 GiB.

Macs that run Gemma 3 12B comfortably

The smallest configuration that handles it is the M3 Pro 18GB (MacBook Pro 14" (M3 Pro)).

Mac	Memory	Bandwidth	Tok/s
M3 Pro 18GB	18 GB	150 GB/s	16
M2 24GB	24 GB	100 GB/s	11
M4 24GB	24 GB	120 GB/s	13
M4 Pro 24GB	24 GB	273 GB/s	29
M1 Max 32GB	32 GB	400 GB/s	43
M2 Max 32GB	32 GB	400 GB/s	43
M4 32GB	32 GB	120 GB/s	13
M3 Max 36GB	36 GB	300 GB/s	32
M4 Max 36GB	36 GB	410 GB/s	44
M4 Pro 48GB	48 GB	273 GB/s	29
M4 Max 48GB	48 GB	546 GB/s	58
M1 Max 64GB	64 GB	400 GB/s	43

Common questions

Can an M1 Pro 16GB Mac run Gemma 3 12B?: Yes. At Q4_K_M it needs about 11.1 GiB of the roughly 13.0 GiB available, generating around 21 tokens per second.
How fast is Gemma 3 12B on an Mac?: Around 21 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 200 GB/s.
How much memory does Gemma 3 12B need?: 7.0 GiB for the weights at Q4_K_M, plus 3.0 GiB for an 8K-token KV cache and about 1.1 GiB of runtime overhead — 11.1 GiB in total.