How fast is DeepSeek R1 on an Mac?

Around 18 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s. Because this is a mixture-of-experts model only 37.44B of its 684.49B parameters are read per token, which is why it is faster than its size suggests.

How much memory does DeepSeek R1 need?

390.5 GiB for the weights at Q4_K_M, plus 0.5 GiB for an 8K-token KV cache and about 20.3 GiB of runtime overhead — 411.3 GiB in total.

Home › DeepSeek R1 › M3 Ultra 512GB

Can an M3 Ultra 512GB Mac run DeepSeek R1?

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

Yes — DeepSeek R1 runs on an M3 Ultra 512GB Mac.

At Q4_K_M it needs about 411.3 GiB, leaving 84.7 GiB spare out of the ~496.0 GiB macOS will let you use. Expect roughly 18 tokens/sec — comfortable for interactive chat.

Weights

390.5 GiB

684.49B params @ Q4_K_M

KV cache

0.5 GiB

8K context, FP16

Overhead

20.3 GiB

runtime + activations

Total needed

411.3 GiB

of ~496.0 GiB usable

Est. speed

18 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	1275.0 GiB	1340.0 GiB	5	No
Q8_0	677.3 GiB	712.5 GiB	10	No
Q6_K	525.9 GiB	553.6 GiB	13	No
Q5_K_M	454.2 GiB	478.3 GiB	15	Yes, but tight
Q4_K_M	390.5 GiB	411.3 GiB	18	Yes, comfortably
Q3_K_M	310.8 GiB	327.6 GiB	22	Yes, comfortably
Q2_K	239.1 GiB	252.3 GiB	29	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	0.1 GiB	410.7 GiB	Yes, comfortably	Yes, comfortably
4K	0.3 GiB	410.9 GiB	Yes, comfortably	Yes, comfortably
8K	0.5 GiB	411.3 GiB	Yes, comfortably	Yes, comfortably
16K	1.1 GiB	412.2 GiB	Yes, comfortably	Yes, comfortably
32K	2.2 GiB	413.8 GiB	Yes, comfortably	Yes, comfortably
128K	8.6 GiB	423.9 GiB	Yes, but tight	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 deepseek-r1

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/DeepSeek-R1-4bit --prompt "Hello"

Macs that run DeepSeek R1 comfortably

The smallest configuration that handles it is the M3 Ultra 512GB (Mac Studio (M3 Ultra)).

Mac	Memory	Bandwidth	Tok/s
M3 Ultra 512GB	512 GB	819 GB/s	18

Common questions

Can an M3 Ultra 512GB Mac run DeepSeek R1?: Yes. At Q4_K_M it needs about 411.3 GiB of the roughly 496.0 GiB available, generating around 18 tokens per second.
How fast is DeepSeek R1 on an Mac?: Around 18 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s. Because this is a mixture-of-experts model only 37.44B of its 684.49B parameters are read per token, which is why it is faster than its size suggests.
How much memory does DeepSeek R1 need?: 390.5 GiB for the weights at Q4_K_M, plus 0.5 GiB for an 8K-token KV cache and about 20.3 GiB of runtime overhead — 411.3 GiB in total.