How fast is gpt-oss-20b on an Mac?

Around 91 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 400 GB/s. Because this is a mixture-of-experts model only 3.6B of its 20.9B parameters are read per token, which is why it is faster than its size suggests.

How much memory does gpt-oss-20b need?

11.9 GiB for the weights at Q4_K_M, plus 0.4 GiB for an 8K-token KV cache and about 1.4 GiB of runtime overhead — 13.7 GiB in total.

Home › gpt-oss-20b › M1 Max 32GB

Can an M1 Max 32GB Mac run gpt-oss-20b?

M1 Max · 32 GB unified memory · 400 GB/s · MacBook Pro 14" (M1 Max), MacBook Pro 16" (M1 Max)

Yes — gpt-oss-20b runs on an M1 Max 32GB Mac.

At Q4_K_M it needs about 13.7 GiB, leaving 13.5 GiB spare out of the ~27.2 GiB macOS will let you use. Expect roughly 91 tokens/sec — faster than you can read.

Weights

11.9 GiB

20.9B params @ Q4_K_M

KV cache

0.4 GiB

8K context, FP16

Overhead

1.4 GiB

runtime + activations

Total needed

13.7 GiB

of ~27.2 GiB usable

Est. speed

91 tok/s

400 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	38.9 GiB	42.1 GiB	28	No
Q8_0	20.7 GiB	22.9 GiB	52	Yes, comfortably
Q6_K	16.1 GiB	18.0 GiB	67	Yes, comfortably
Q5_K_M	13.9 GiB	15.7 GiB	78	Yes, comfortably
Q4_K_M	11.9 GiB	13.7 GiB	91	Yes, comfortably
Q3_K_M	9.5 GiB	11.1 GiB	114	Yes, comfortably
Q2_K	7.3 GiB	8.8 GiB	148	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	13.2 GiB	Yes, comfortably	Yes, comfortably
4K	0.2 GiB	13.4 GiB	Yes, comfortably	Yes, comfortably
8K	0.4 GiB	13.7 GiB	Yes, comfortably	Yes, comfortably
16K	0.8 GiB	14.4 GiB	Yes, comfortably	Yes, comfortably
32K	1.5 GiB	15.7 GiB	Yes, comfortably	Yes, comfortably
128K	6.0 GiB	23.8 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 gpt-oss:20b

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/gpt-oss-20b-4bit --prompt "Hello"

Macs that run gpt-oss-20b comfortably

The smallest configuration that handles it is the M2 24GB (MacBook Air 15" (M2)).

Mac	Memory	Bandwidth	Tok/s
M2 24GB	24 GB	100 GB/s	23
M4 24GB	24 GB	120 GB/s	27
M4 Pro 24GB	24 GB	273 GB/s	62
M1 Max 32GB	32 GB	400 GB/s	91
M2 Max 32GB	32 GB	400 GB/s	91
M4 32GB	32 GB	120 GB/s	27
M3 Max 36GB	36 GB	300 GB/s	68
M4 Max 36GB	36 GB	410 GB/s	93
M4 Pro 48GB	48 GB	273 GB/s	62
M4 Max 48GB	48 GB	546 GB/s	124
M1 Max 64GB	64 GB	400 GB/s	91
M2 Ultra 64GB	64 GB	800 GB/s	181

Common questions

Can an M1 Max 32GB Mac run gpt-oss-20b?: Yes. At Q4_K_M it needs about 13.7 GiB of the roughly 27.2 GiB available, generating around 91 tokens per second.
How fast is gpt-oss-20b on an Mac?: Around 91 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 400 GB/s. Because this is a mixture-of-experts model only 3.6B of its 20.9B parameters are read per token, which is why it is faster than its size suggests.
How much memory does gpt-oss-20b need?: 11.9 GiB for the weights at Q4_K_M, plus 0.4 GiB for an 8K-token KV cache and about 1.4 GiB of runtime overhead — 13.7 GiB in total.