Can It Run?
HomeModels › DeepSeek R1

DeepSeek R1 on a Mac

DeepSeek · 684.49B parameters (37.44B active — mixture of experts) · mit · released 2025-01

Weights @ Q4_K_M
390.5 GiB
the usual download size
KV cache
69 KB
per token, FP16
Max context
128K
tokens
Smallest Mac
512 GB
M3 Ultra
Short answerA M3 Ultra 512GB is the smallest Apple Silicon machine that loads DeepSeek R1 at Q4_K_M with an 8K context, at roughly 18 tokens/sec. Mac Studio (M3 Ultra) is the cheapest way to get that configuration.

Every Mac, ranked

MacMemoryNeedsTok/sVerdict
M1 8GB8 GB411.3 GiB1No
M1 16GB16 GB411.3 GiB1No
M2 16GB16 GB411.3 GiB2No
M4 16GB16 GB411.3 GiB3No
M1 Pro 16GB16 GB411.3 GiB4No
M3 Pro 18GB18 GB411.3 GiB3No
M2 24GB24 GB411.3 GiB2No
M4 24GB24 GB411.3 GiB3No
M4 Pro 24GB24 GB411.3 GiB6No
M4 32GB32 GB411.3 GiB3No
M1 Max 32GB32 GB411.3 GiB9No
M2 Max 32GB32 GB411.3 GiB9No
M3 Max 36GB36 GB411.3 GiB7No
M4 Max 36GB36 GB411.3 GiB9No
M4 Pro 48GB48 GB411.3 GiB6No
M4 Max 48GB48 GB411.3 GiB12No
M4 Pro 64GB64 GB411.3 GiB6No
M1 Max 64GB64 GB411.3 GiB9No
M4 Max 64GB64 GB411.3 GiB12No
M2 Ultra 64GB64 GB411.3 GiB17No
M2 Max 96GB96 GB411.3 GiB9No
M3 Ultra 96GB96 GB411.3 GiB18No
M3 Max 128GB128 GB411.3 GiB9No
M4 Max 128GB128 GB411.3 GiB12No
M2 Ultra 192GB192 GB411.3 GiB17No
M3 Ultra 256GB256 GB411.3 GiB18No
M3 Ultra 512GB512 GB411.3 GiB18Yes, comfortably

Download size by quantisation

Weight memory scales linearly with bits per parameter. Everything below assumes an 8K context on top.

QuantisationWeightsTotal @ 8KNotes
FP161275.0 GiB1340.0 GiBFull precision. Reference quality, twice the memory of Q8.
Q8_0677.3 GiB712.5 GiBIndistinguishable from FP16 in practice, at half the size.
Q6_K525.9 GiB553.6 GiBNear-lossless. A good stop when you have memory to spare.
Q5_K_M454.2 GiB478.3 GiBSlightly better than Q4_K_M, noticeably bigger.
Q4_K_M390.5 GiB411.3 GiBThe default. Best quality-per-gigabyte for most people.
Q3_K_M310.8 GiB327.6 GiBVisible quality loss. Use to squeeze one size class up.
Q2_K239.1 GiB252.3 GiBLast resort. Often worse than a smaller model at Q4.
Mixture of expertsDeepSeek R1 holds 684.49B parameters in memory but only reads about 37.44B per token. You pay the full memory cost of a 684.49B model and get roughly the speed of a 37.44B one — an excellent trade if you have the RAM.