How fast is Phi-4 14B on an Mac?

Around 73 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s.

How much memory does Phi-4 14B need?

8.4 GiB for the weights at Q4_K_M, plus 1.6 GiB for an 8K-token KV cache and about 1.2 GiB of runtime overhead — 11.2 GiB in total.

Home › Phi-4 14B › M3 Ultra 512GB

Can an M3 Ultra 512GB Mac run Phi-4 14B?

M3 Ultra · 512 GB unified memory · 819 GB/s · Mac Studio (M3 Ultra)

Yes — Phi-4 14B runs on an M3 Ultra 512GB Mac.

At Q4_K_M it needs about 11.2 GiB, leaving 484.8 GiB spare out of the ~496.0 GiB macOS will let you use. Expect roughly 73 tokens/sec — faster than you can read.

Weights

8.4 GiB

14.7B params @ Q4_K_M

KV cache

1.6 GiB

8K context, FP16

Overhead

1.2 GiB

runtime + activations

Total needed

11.2 GiB

of ~496.0 GiB usable

Est. speed

73 tok/s

819 GB/s bandwidth

Max context

16K 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	27.4 GiB	31.1 GiB	22	Yes, comfortably
Q8_0	14.5 GiB	17.6 GiB	42	Yes, comfortably
Q6_K	11.3 GiB	14.2 GiB	54	Yes, comfortably
Q5_K_M	9.8 GiB	12.6 GiB	63	Yes, comfortably
Q4_K_M	8.4 GiB	11.2 GiB	73	Yes, comfortably
Q3_K_M	6.7 GiB	9.4 GiB	91	Yes, comfortably
Q2_K	5.1 GiB	7.8 GiB	119	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	0.4 GiB	9.8 GiB	Yes, comfortably	Yes, comfortably
4K	0.8 GiB	10.2 GiB	Yes, comfortably	Yes, comfortably
8K	1.6 GiB	11.2 GiB	Yes, comfortably	Yes, comfortably
16K	3.1 GiB	13.0 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 phi4

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/Phi-4-14B-4bit --prompt "Hello"

Macs that run Phi-4 14B 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	13
M2 24GB	24 GB	100 GB/s	9
M4 24GB	24 GB	120 GB/s	11
M4 Pro 24GB	24 GB	273 GB/s	24
M1 Max 32GB	32 GB	400 GB/s	36
M2 Max 32GB	32 GB	400 GB/s	36
M4 32GB	32 GB	120 GB/s	11
M3 Max 36GB	36 GB	300 GB/s	27
M4 Max 36GB	36 GB	410 GB/s	36
M4 Pro 48GB	48 GB	273 GB/s	24
M4 Max 48GB	48 GB	546 GB/s	49
M1 Max 64GB	64 GB	400 GB/s	36

Common questions

Can an M3 Ultra 512GB Mac run Phi-4 14B?: Yes. At Q4_K_M it needs about 11.2 GiB of the roughly 496.0 GiB available, generating around 73 tokens per second.
How fast is Phi-4 14B on an Mac?: Around 73 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s.
How much memory does Phi-4 14B need?: 8.4 GiB for the weights at Q4_K_M, plus 1.6 GiB for an 8K-token KV cache and about 1.2 GiB of runtime overhead — 11.2 GiB in total.