Skip to content
mlx.app

GPT-OSS 120B

OpenAI · 116.8B (5.1B active, MoE) · up to 128K context · apache-2.0

OVER BY 27.9 GB

This is the model that made a 64GB M-series Mac a genuinely useful frontier-model machine, since only 5.1B parameters activate per token despite the 117B total. On a 128GB Mac Studio it runs quickly enough for interactive use, which is unusual for a model this large.

On a M4 Pro with 48 GB

macOS9.6 GBweights65.7 GBKV cache0.6 GBover27.9 GB
Weights
65.7 GB
KV cache
0.6 GB
Usable RAM
38.4 GB
Headroom
0.0 GB
Generation
74.2 tok/sest.
Prompt processing
177 tok/sest.
First token (1K prompt)
5.8 sest.

Decode speed is bandwidth ÷ active weight bytes, at 273 GB/s and 78% efficiency. These are modelled figures, not measurements.

Every quantization, on your Mac

QuantWeights+ KVVerdictSpeed est.Repo
q465.7 GB66.3 GBOVER BY 27.9 GB74.2 tok/sgpt-oss-120b-MXFP4-Q4
q8124.1 GB124.7 GBOVER BY 86.3 GB39.3 tok/sgpt-oss-120b-MXFP4-Q8

We never host weights. Every link goes to Hugging Face.

Run it

Chat
mlx_lm.chat --model mlx-community/gpt-oss-120b-MXFP4-Q4
Serve an OpenAI-compatible endpoint
mlx_lm.server --model mlx-community/gpt-oss-120b-MXFP4-Q4 --port 8080

How far can you push the context

2K tokensOVER BY 27.5 GB

cache 0.2 GB

8K tokensOVER BY 27.9 GB

cache 0.6 GB

32K tokensOVER BY 29.7 GB

cache 2.4 GB

128K tokensOVER BY 37.0 GB

cache 9.7 GB

Which Macs run this

ChipSmallest RAM that fitsSpeed est.
M1no configuration
M1 Prono configuration
M1 Maxno configuration
M1 Ultra128 GB226 tok/s
M2no configuration
M2 Prono configuration
M2 Max96 GB112 tok/s
M2 Ultra128 GB229 tok/s
M3no configuration
M3 Prono configuration
M3 Max (14-core CPU)no configuration
M3 Max (16-core CPU)96 GB112 tok/s
M3 Ultra96 GB237 tok/s
M4no configuration
M4 Prono configuration
M4 Max (14-core CPU)no configuration
M4 Max (16-core CPU)128 GB156 tok/s
M5no configuration
M5 Prounverifiedno configuration
M5 Maxunverified128 GB178 tok/s