How fast is Qwen2.5 72B on an Mac?

Around 14 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 800 GB/s.

How much memory does Qwen2.5 72B need?

41.5 GiB for the weights at Q4_K_M, plus 2.5 GiB for an 8K-token KV cache and about 2.9 GiB of runtime overhead — 46.8 GiB in total.

Home › Qwen2.5 72B › M2 Ultra 64GB

Can an M2 Ultra 64GB Mac run Qwen2.5 72B?

M2 Ultra · 64 GB unified memory · 800 GB/s · Mac Studio (M2 Ultra), Mac Pro (M2 Ultra)

Just barely — and you will feel it.

Qwen2.5 72B at Q4_K_M needs about 46.8 GiB against roughly 54.4 GiB usable — that is 86% of your budget. It will load, but close your browser first, keep the context short, and expect memory pressure. Around 14 tokens/sec.

Weights

41.5 GiB

72.7B params @ Q4_K_M

KV cache

2.5 GiB

8K context, FP16

Overhead

2.9 GiB

runtime + activations

Total needed

46.8 GiB

of ~54.4 GiB usable

Est. speed

14 tok/s

800 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	135.4 GiB	145.5 GiB	4	No
Q8_0	71.9 GiB	78.8 GiB	8	No
Q6_K	55.9 GiB	62.0 GiB	11	No
Q5_K_M	48.2 GiB	54.0 GiB	12	Yes, but tight
Q4_K_M	41.5 GiB	46.8 GiB	14	Yes, but tight
Q3_K_M	33.0 GiB	38.0 GiB	18	Yes, comfortably
Q2_K	25.4 GiB	30.0 GiB	23	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	0.6 GiB	44.7 GiB	Yes, comfortably	Yes, comfortably
4K	1.3 GiB	45.4 GiB	Yes, comfortably	Yes, comfortably
8K	2.5 GiB	46.8 GiB	Yes, but tight	Yes, comfortably
16K	5.0 GiB	49.6 GiB	Yes, but tight	Yes, but tight
32K	10.0 GiB	55.2 GiB	No	Yes, but tight
128K	40.0 GiB	88.8 GiB	No	No

What to run instead on an M2 Ultra 64GB 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
DeepSeek-R1-Distill-Llama-70B	70.6B	45.6 GiB	15	Yes, comfortably
Llama 3.3 70B	70.6B	45.6 GiB	15	Yes, comfortably
Mixtral 8x7B	46.7B	29.8 GiB	51	Yes, comfortably
DeepSeek-R1-Distill-Qwen-32B	32.8B	22.4 GiB	32	Yes, comfortably
Qwen2.5 32B	32.8B	22.4 GiB	32	Yes, comfortably
Qwen2.5-Coder 32B	32.8B	22.4 GiB	32	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 qwen2.5:72b

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/Qwen2-5-72B-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=55705. This resets on reboot. Quantising the KV cache (--kv-cache-type q8_0 in llama.cpp) buys back 1.2 GiB.

Macs that run Qwen2.5 72B comfortably

The smallest configuration that handles it is the M2 Max 96GB (MacBook Pro 16" (M2 Max)).

Mac	Memory	Bandwidth	Tok/s
M2 Max 96GB	96 GB	400 GB/s	7
M3 Ultra 96GB	96 GB	819 GB/s	15
M3 Max 128GB	128 GB	400 GB/s	7
M4 Max 128GB	128 GB	546 GB/s	10
M2 Ultra 192GB	192 GB	800 GB/s	14
M3 Ultra 256GB	256 GB	819 GB/s	15
M3 Ultra 512GB	512 GB	819 GB/s	15

Common questions

Can an M2 Ultra 64GB Mac run Qwen2.5 72B?: Yes. At Q4_K_M it needs about 46.8 GiB of the roughly 54.4 GiB available, generating around 14 tokens per second.
How fast is Qwen2.5 72B on an Mac?: Around 14 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 800 GB/s.
How much memory does Qwen2.5 72B need?: 41.5 GiB for the weights at Q4_K_M, plus 2.5 GiB for an 8K-token KV cache and about 2.9 GiB of runtime overhead — 46.8 GiB in total.