How fast is Qwen2.5 32B on an Mac?

Around 33 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s.

How much memory does Qwen2.5 32B need?

18.7 GiB for the weights at Q4_K_M, plus 2.0 GiB for an 8K-token KV cache and about 1.7 GiB of runtime overhead — 22.4 GiB in total.

Home › Qwen2.5 32B › M3 Ultra 256GB

Can an M3 Ultra 256GB Mac run Qwen2.5 32B?

M3 Ultra · 256 GB unified memory · 819 GB/s · Mac Studio (M3 Ultra)

Yes — Qwen2.5 32B runs on an M3 Ultra 256GB Mac.

At Q4_K_M it needs about 22.4 GiB, leaving 217.6 GiB spare out of the ~240.0 GiB macOS will let you use. Expect roughly 33 tokens/sec — comfortable for interactive chat.

Weights

18.7 GiB

32.8B params @ Q4_K_M

KV cache

2.0 GiB

8K context, FP16

Overhead

1.7 GiB

runtime + activations

Total needed

22.4 GiB

of ~240.0 GiB usable

Est. speed

33 tok/s

819 GB/s bandwidth

Max context

128K 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	61.1 GiB	66.9 GiB	10	Yes, comfortably
Q8_0	32.5 GiB	36.9 GiB	19	Yes, comfortably
Q6_K	25.2 GiB	29.3 GiB	24	Yes, comfortably
Q5_K_M	21.8 GiB	25.7 GiB	28	Yes, comfortably
Q4_K_M	18.7 GiB	22.4 GiB	33	Yes, comfortably
Q3_K_M	14.9 GiB	18.4 GiB	41	Yes, comfortably
Q2_K	11.5 GiB	14.8 GiB	53	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	0.5 GiB	20.7 GiB	Yes, comfortably	Yes, comfortably
4K	1.0 GiB	21.3 GiB	Yes, comfortably	Yes, comfortably
8K	2.0 GiB	22.4 GiB	Yes, comfortably	Yes, comfortably
16K	4.0 GiB	24.7 GiB	Yes, comfortably	Yes, comfortably
32K	8.0 GiB	29.3 GiB	Yes, comfortably	Yes, comfortably
128K	32.0 GiB	56.9 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 qwen2.5:32b

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/Qwen2-5-32B-4bit --prompt "Hello"

Macs that run Qwen2.5 32B comfortably

The smallest configuration that handles it is the M1 Max 32GB (MacBook Pro 14" (M1 Max)).

Mac	Memory	Bandwidth	Tok/s
M1 Max 32GB	32 GB	400 GB/s	16
M2 Max 32GB	32 GB	400 GB/s	16
M4 32GB	32 GB	120 GB/s	5
M3 Max 36GB	36 GB	300 GB/s	12
M4 Max 36GB	36 GB	410 GB/s	16
M4 Pro 48GB	48 GB	273 GB/s	11
M4 Max 48GB	48 GB	546 GB/s	22
M1 Max 64GB	64 GB	400 GB/s	16
M2 Ultra 64GB	64 GB	800 GB/s	32
M4 Pro 64GB	64 GB	273 GB/s	11
M4 Max 64GB	64 GB	546 GB/s	22
M2 Max 96GB	96 GB	400 GB/s	16

Common questions

Can an M3 Ultra 256GB Mac run Qwen2.5 32B?: Yes. At Q4_K_M it needs about 22.4 GiB of the roughly 240.0 GiB available, generating around 33 tokens per second.
How fast is Qwen2.5 32B on an Mac?: Around 33 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s.
How much memory does Qwen2.5 32B need?: 18.7 GiB for the weights at Q4_K_M, plus 2.0 GiB for an 8K-token KV cache and about 1.7 GiB of runtime overhead — 22.4 GiB in total.