How fast is DeepSeek V4 Flash 0731 on an Mac?

Around 4 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s.

How much memory does DeepSeek V4 Flash 0731 need?

173.5 GiB for the weights at Q4_K_M, plus 0.7 GiB for an 8K-token KV cache and about 9.5 GiB of runtime overhead — 183.7 GiB in total.

Home › DeepSeek V4 Flash 0731 › M3 Ultra 512GB

Can an M3 Ultra 512GB Mac run DeepSeek V4 Flash 0731?

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

Yes — DeepSeek V4 Flash 0731 runs on an M3 Ultra 512GB Mac.

At Q4_K_M it needs about 183.7 GiB, leaving 312.3 GiB spare out of the ~496.0 GiB macOS will let you use. Expect roughly 4 tokens/sec — too slow for interactive use.

Weights

173.5 GiB

304.18B params @ Q4_K_M

KV cache

0.7 GiB

8K context, FP16

Overhead

9.5 GiB

runtime + activations

Total needed

183.7 GiB

of ~496.0 GiB usable

Est. speed

4 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	566.6 GiB	596.4 GiB	1	No
Q8_0	301.0 GiB	317.5 GiB	2	Yes, comfortably
Q6_K	233.7 GiB	246.9 GiB	3	Yes, comfortably
Q5_K_M	201.8 GiB	213.4 GiB	3	Yes, comfortably
Q4_K_M	173.5 GiB	183.7 GiB	4	Yes, comfortably
Q3_K_M	138.1 GiB	146.5 GiB	4	Yes, comfortably
Q2_K	106.2 GiB	113.0 GiB	6	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	182.9 GiB	Yes, comfortably	Yes, comfortably
4K	0.3 GiB	183.2 GiB	Yes, comfortably	Yes, comfortably
8K	0.7 GiB	183.7 GiB	Yes, comfortably	Yes, comfortably
16K	1.3 GiB	184.6 GiB	Yes, comfortably	Yes, comfortably
32K	2.7 GiB	186.6 GiB	Yes, comfortably	Yes, comfortably
128K	10.8 GiB	198.2 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 deepseek-v4-flash-0731

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-V4-Flash-0731-4bit --prompt "Hello"

Macs that run DeepSeek V4 Flash 0731 comfortably

The smallest configuration that handles it is the M3 Ultra 256GB (Mac Studio (M3 Ultra)).

Mac	Memory	Bandwidth	Tok/s
M3 Ultra 256GB	256 GB	819 GB/s	4
M3 Ultra 512GB	512 GB	819 GB/s	4

Common questions

Can an M3 Ultra 512GB Mac run DeepSeek V4 Flash 0731?: Yes. At Q4_K_M it needs about 183.7 GiB of the roughly 496.0 GiB available, generating around 4 tokens per second.
How fast is DeepSeek V4 Flash 0731 on an Mac?: Around 4 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s.
How much memory does DeepSeek V4 Flash 0731 need?: 173.5 GiB for the weights at Q4_K_M, plus 0.7 GiB for an 8K-token KV cache and about 9.5 GiB of runtime overhead — 183.7 GiB in total.