How fast is DeepSeek V4 Flash 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 need?

166.0 GiB for the weights at Q4_K_M, plus 0.7 GiB for an 8K-token KV cache and about 9.1 GiB of runtime overhead — 175.7 GiB in total.

Home › DeepSeek V4 Flash › M3 Ultra 96GB

Can an M3 Ultra 96GB Mac run DeepSeek V4 Flash?

M3 Ultra · 96 GB unified memory · 819 GB/s · Mac Studio (M3 Ultra)

No — not at Q4_K_M.

DeepSeek V4 Flash needs about 175.7 GiB but an M3 Ultra 96GB Mac has only ~81.6 GiB available for a model — you are 94.1 GiB short. macOS will swap to SSD and generation will crawl.

Weights

166.0 GiB

290.94B params @ Q4_K_M

KV cache

0.7 GiB

8K context, FP16

Overhead

9.1 GiB

runtime + activations

Total needed

175.7 GiB

of ~81.6 GiB usable

Est. speed

4 tok/s

819 GB/s bandwidth

Max context

n/a

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	541.9 GiB	570.5 GiB	1	No
Q8_0	287.9 GiB	303.8 GiB	2	No
Q6_K	223.5 GiB	236.2 GiB	3	No
Q5_K_M	193.1 GiB	204.2 GiB	3	No
Q4_K_M	166.0 GiB	175.7 GiB	4	No
Q3_K_M	132.1 GiB	140.2 GiB	5	No
Q2_K	101.6 GiB	108.2 GiB	6	No

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	175.0 GiB	No	No
4K	0.3 GiB	175.2 GiB	No	No
8K	0.7 GiB	175.7 GiB	No	No
16K	1.3 GiB	176.7 GiB	No	No
32K	2.7 GiB	178.6 GiB	No	No
128K	10.8 GiB	190.3 GiB	No	No

What to run instead on an M3 Ultra 96GB 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
Qwen2.5 72B	72.7B	46.8 GiB	15	Yes, comfortably
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	52	Yes, comfortably
DeepSeek-R1-Distill-Qwen-32B	32.8B	22.4 GiB	33	Yes, comfortably
Qwen2.5 32B	32.8B	22.4 GiB	33	Yes, comfortably

Macs that run DeepSeek V4 Flash 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 96GB Mac run DeepSeek V4 Flash?: No. At Q4_K_M it needs about 175.7 GiB, and an M3 Ultra 96GB Mac has roughly 81.6 GiB available for a model after macOS takes its share.
How fast is DeepSeek V4 Flash 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 need?: 166.0 GiB for the weights at Q4_K_M, plus 0.7 GiB for an 8K-token KV cache and about 9.1 GiB of runtime overhead — 175.7 GiB in total.