How fast is gpt-oss-120b on an Mac?

Around 131 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s. Because this is a mixture-of-experts model only 5.1B of its 116.8B parameters are read per token, which is why it is faster than its size suggests.

How much memory does gpt-oss-120b need?

66.6 GiB for the weights at Q4_K_M, plus 0.6 GiB for an 8K-token KV cache and about 4.1 GiB of runtime overhead — 71.3 GiB in total.

Home › gpt-oss-120b › M3 Ultra 96GB

Can an M3 Ultra 96GB Mac run gpt-oss-120b?

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

Just barely — and you will feel it.

gpt-oss-120b at Q4_K_M needs about 71.3 GiB against roughly 81.6 GiB usable — that is 87% of your budget. It will load, but close your browser first, keep the context short, and expect memory pressure. Around 131 tokens/sec.

Weights

66.6 GiB

116.8B params @ Q4_K_M

KV cache

0.6 GiB

8K context, FP16

Overhead

4.1 GiB

runtime + activations

Total needed

71.3 GiB

of ~81.6 GiB usable

Est. speed

131 tok/s

819 GB/s bandwidth

Max context

64K 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	217.6 GiB	229.8 GiB	40	No
Q8_0	115.6 GiB	122.7 GiB	76	No
Q6_K	89.7 GiB	95.6 GiB	97	No
Q5_K_M	77.5 GiB	82.7 GiB	113	No
Q4_K_M	66.6 GiB	71.3 GiB	131	Yes, but tight
Q3_K_M	53.0 GiB	57.0 GiB	165	Yes, comfortably
Q2_K	40.8 GiB	44.2 GiB	214	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	70.7 GiB	Yes, but tight	Yes, but tight
4K	0.3 GiB	70.9 GiB	Yes, but tight	Yes, but tight
8K	0.6 GiB	71.3 GiB	Yes, but tight	Yes, but tight
16K	1.1 GiB	72.2 GiB	Yes, but tight	Yes, but tight
32K	2.3 GiB	73.9 GiB	Yes, but tight	Yes, but tight
128K	9.0 GiB	84.3 GiB	No	Yes, but tight

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

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:120b

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-120b-4bit --prompt "Hello"

You are close to the limitIf it stutters or gets killed, raise the Metal memory cap before loading: sudo sysctl iogpu.wired_limit_mb=83558. This resets on reboot. Quantising the KV cache (--kv-cache-type q8_0 in llama.cpp) buys back 0.3 GiB.

Macs that run gpt-oss-120b comfortably

The smallest configuration that handles it is the M3 Max 128GB (MacBook Pro 16" (M3 Max)).

Mac	Memory	Bandwidth	Tok/s
M3 Max 128GB	128 GB	400 GB/s	64
M4 Max 128GB	128 GB	546 GB/s	87
M2 Ultra 192GB	192 GB	800 GB/s	128
M3 Ultra 256GB	256 GB	819 GB/s	131
M3 Ultra 512GB	512 GB	819 GB/s	131

Common questions

Can an M3 Ultra 96GB Mac run gpt-oss-120b?: Yes. At Q4_K_M it needs about 71.3 GiB of the roughly 81.6 GiB available, generating around 131 tokens per second.
How fast is gpt-oss-120b on an Mac?: Around 131 tokens per second. Token generation on Apple Silicon is limited by memory bandwidth, and this machine has 819 GB/s. Because this is a mixture-of-experts model only 5.1B of its 116.8B parameters are read per token, which is why it is faster than its size suggests.
How much memory does gpt-oss-120b need?: 66.6 GiB for the weights at Q4_K_M, plus 0.6 GiB for an 8K-token KV cache and about 4.1 GiB of runtime overhead — 71.3 GiB in total.