Can It Run?
HomeModels › gpt-oss-20b

gpt-oss-20b on a Mac

OpenAI · 20.9B parameters (3.6B active — mixture of experts) · Apache 2.0 · released 2025-08

Weights @ Q4_K_M
11.9 GiB
the usual download size
KV cache
48 KB
per token, FP16
Max context
128K
tokens
Smallest Mac
18 GB
M3 Pro
Short answerA M3 Pro 18GB is the smallest Apple Silicon machine that loads gpt-oss-20b at Q4_K_M with an 8K context, at roughly 34 tokens/sec. MacBook Pro 14" (M3 Pro) is the cheapest way to get that configuration.

Every Mac, ranked

MacMemoryNeedsTok/sVerdict
M1 8GB8 GB13.7 GiB15No
M1 16GB16 GB13.7 GiB15No
M2 16GB16 GB13.7 GiB23No
M4 16GB16 GB13.7 GiB27No
M1 Pro 16GB16 GB13.7 GiB45No
M3 Pro 18GB18 GB13.7 GiB34Yes, but tight
M2 24GB24 GB13.7 GiB23Yes, comfortably
M4 24GB24 GB13.7 GiB27Yes, comfortably
M4 Pro 24GB24 GB13.7 GiB62Yes, comfortably
M4 32GB32 GB13.7 GiB27Yes, comfortably
M1 Max 32GB32 GB13.7 GiB91Yes, comfortably
M2 Max 32GB32 GB13.7 GiB91Yes, comfortably
M3 Max 36GB36 GB13.7 GiB68Yes, comfortably
M4 Max 36GB36 GB13.7 GiB93Yes, comfortably
M4 Pro 48GB48 GB13.7 GiB62Yes, comfortably
M4 Max 48GB48 GB13.7 GiB124Yes, comfortably
M4 Pro 64GB64 GB13.7 GiB62Yes, comfortably
M1 Max 64GB64 GB13.7 GiB91Yes, comfortably
M4 Max 64GB64 GB13.7 GiB124Yes, comfortably
M2 Ultra 64GB64 GB13.7 GiB181Yes, comfortably
M2 Max 96GB96 GB13.7 GiB91Yes, comfortably
M3 Ultra 96GB96 GB13.7 GiB186Yes, comfortably
M3 Max 128GB128 GB13.7 GiB91Yes, comfortably
M4 Max 128GB128 GB13.7 GiB124Yes, comfortably
M2 Ultra 192GB192 GB13.7 GiB181Yes, comfortably
M3 Ultra 256GB256 GB13.7 GiB186Yes, comfortably
M3 Ultra 512GB512 GB13.7 GiB186Yes, comfortably

Download size by quantisation

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

QuantisationWeightsTotal @ 8KNotes
FP1638.9 GiB42.1 GiBFull precision. Reference quality, twice the memory of Q8.
Q8_020.7 GiB22.9 GiBIndistinguishable from FP16 in practice, at half the size.
Q6_K16.1 GiB18.0 GiBNear-lossless. A good stop when you have memory to spare.
Q5_K_M13.9 GiB15.7 GiBSlightly better than Q4_K_M, noticeably bigger.
Q4_K_M11.9 GiB13.7 GiBThe default. Best quality-per-gigabyte for most people.
Q3_K_M9.5 GiB11.1 GiBVisible quality loss. Use to squeeze one size class up.
Q2_K7.3 GiB8.8 GiBLast resort. Often worse than a smaller model at Q4.
Mixture of expertsgpt-oss-20b holds 20.9B parameters in memory but only reads about 3.6B per token. You pay the full memory cost of a 20.9B model and get roughly the speed of a 3.6B one — an excellent trade if you have the RAM.