gpt-oss-120b is a 120-billion-parameter model you can actually run at home — if you know the MoE trick

The short answer

gpt-oss-120b is a Mixture-of-Experts model: 120B total parameters, ~5B active per token, shipping in MXFP4 at ~60GB of weights. It needs enough memory to hold all 120B parameters (~60GB) but only…

120B total, ~5B active — MoE means low compute per token, but all params must be in memory.

~60GB in MXFP4 — the weights fit a 64GB+ unified-memory machine or a big multi-GPU setup.

24GB GPU + offload: llama.cpp's --n-cpu-moe pushes experts to system RAM to run it on less VRAM.

Best homes: a 96GB+ Strix Halo or Mac, or a dual-GPU rig.

Not a datacenter model — the MoE design makes 120B genuinely runnable at home.