How fast is Qwen3 30B-A3B on an Mac?

Around 68 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 273 GB/s. Because this is a mixture-of-experts model only 3.3B of its 30.5B parameters are read per token, which is why it is faster than its size suggests.

How much memory does Qwen3 30B-A3B need?

17.4 GiB for the weights at Q4_K_M, plus 0.8 GiB for an 8K-token KV cache and about 1.7 GiB of runtime overhead — 19.8 GiB in total.

Home › Qwen3 30B-A3B › M4 Pro 24GB

Can an M4 Pro 24GB Mac run Qwen3 30B-A3B?

M4 Pro · 24 GB unified memory · 273 GB/s · Mac mini (M4 Pro), MacBook Pro 14" (M4 Pro)

Just barely — and you will feel it.

Qwen3 30B-A3B at Q4_K_M needs about 19.8 GiB against roughly 20.4 GiB usable — that is 97% of your budget. It will load, but close your browser first, keep the context short, and expect memory pressure. Around 68 tokens/sec.

Weights

17.4 GiB

30.5B params @ Q4_K_M

KV cache

0.8 GiB

8K context, FP16

Overhead

1.7 GiB

runtime + activations

Total needed

19.8 GiB

of ~20.4 GiB usable

Est. speed

68 tok/s

273 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	56.8 GiB	61.2 GiB	21	No
Q8_0	30.2 GiB	33.2 GiB	39	No
Q6_K	23.4 GiB	26.2 GiB	50	No
Q5_K_M	20.2 GiB	22.8 GiB	58	No
Q4_K_M	17.4 GiB	19.8 GiB	68	Yes, but tight
Q3_K_M	13.8 GiB	16.1 GiB	85	Yes, comfortably
Q2_K	10.7 GiB	12.7 GiB	110	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	0.2 GiB	19.0 GiB	Yes, but tight	Yes, but tight
4K	0.4 GiB	19.3 GiB	Yes, but tight	Yes, but tight
8K	0.8 GiB	19.8 GiB	Yes, but tight	Yes, but tight
16K	1.5 GiB	20.9 GiB	No	Yes, but tight
32K	3.0 GiB	23.0 GiB	No	No
128K	12.0 GiB	35.6 GiB	No	No

What to run instead on an M4 Pro 24GB 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
Mistral Small 3 24B	23.6B	16.2 GiB	15	Yes, comfortably
gpt-oss-20b	20.9B	13.7 GiB	62	Yes, comfortably
DeepSeek-R1-Distill-Qwen-14B	14.8B	11.2 GiB	24	Yes, comfortably
Qwen2.5 14B	14.8B	11.2 GiB	24	Yes, comfortably
Qwen2.5-Coder 14B	14.8B	11.2 GiB	24	Yes, comfortably
Qwen3 14B	14.8B	10.9 GiB	24	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 qwen3:30b-a3b

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/Qwen3-30B-A3B-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=20889. This resets on reboot. Quantising the KV cache (--kv-cache-type q8_0 in llama.cpp) buys back 0.4 GiB.

Macs that run Qwen3 30B-A3B 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	99
M2 Max 32GB	32 GB	400 GB/s	99
M4 32GB	32 GB	120 GB/s	30
M3 Max 36GB	36 GB	300 GB/s	74
M4 Max 36GB	36 GB	410 GB/s	101
M4 Pro 48GB	48 GB	273 GB/s	68
M4 Max 48GB	48 GB	546 GB/s	135
M1 Max 64GB	64 GB	400 GB/s	99
M2 Ultra 64GB	64 GB	800 GB/s	198
M4 Pro 64GB	64 GB	273 GB/s	68
M4 Max 64GB	64 GB	546 GB/s	135
M2 Max 96GB	96 GB	400 GB/s	99

Common questions

Can an M4 Pro 24GB Mac run Qwen3 30B-A3B?: Yes. At Q4_K_M it needs about 19.8 GiB of the roughly 20.4 GiB available, generating around 68 tokens per second.
How fast is Qwen3 30B-A3B on an Mac?: Around 68 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 273 GB/s. Because this is a mixture-of-experts model only 3.3B of its 30.5B parameters are read per token, which is why it is faster than its size suggests.
How much memory does Qwen3 30B-A3B need?: 17.4 GiB for the weights at Q4_K_M, plus 0.8 GiB for an 8K-token KV cache and about 1.7 GiB of runtime overhead — 19.8 GiB in total.