Can I run GPT-OSS 120B?

GPT-OSS 120B is OpenAI's big open-weight release for people who want a frontier-style chat and reasoning model running entirely on their own hardware. It is a mixture-of-experts design, so even though the full model is 116.8B parameters, only about 5.1B are active per token. That is the trick that keeps it fast for its size, but do not let it fool you on memory: you still have to hold the whole thing in RAM. At a 4-bit quant that is roughly 71 GB, and you want around 96 GB of system memory to run it comfortably. This is not a laptop-GPU model.

In practice that means an RTX 3060 or even a 4090 simply will not fit it, so the realistic homes are a high-memory Apple Silicon machine or a big-RAM CPU box. On an M-series Max you can expect around 113 tokens per second, which feels genuinely snappy for a model this large. On CPU with DDR5 you drop to about 16 tokens per second, usable for batch work but slow for interactive chat. The 128K context window is generous, but filling it pushes total memory to roughly 127 GB at max context, so plan for headroom rather than treating that ceiling as your normal working size.

Against its peers, GPT-OSS 120B leans hard toward chat and reasoning rather than coding or vision, so something like Mistral Small 4 119B generally gives you more range if you need code or image input in the same footprint. Its own little sibling, GPT-OSS 20B, is the one to reach for when 96 GB of RAM is out of the question. The standout here is the license: Apache 2.0 means you can use it commercially and in production with no provider strings attached, which is rare at this capability tier and a real reason to pick it.