Only with CPU offload

gpt-oss 120B at MXFP4 needs 61.4 GB but only 14.4 GB is addressable, so about 78% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 6.7 tokens per second — usable for batch work, painful for chat.

22% on GPU 8K context MXFP4 · 60.5 GB Apache 2.0 Released Aug 2025 Native MXFP4

Designed to land on one 80 GB card. Only ~5B parameters are active per token.

What hardware do I need for gpt-oss 120B? →

Fits instead: gpt-oss 20B (11.6 GB)

The VRAM budget

weights 60.5 GB

Weights 60.5 GB KV cache @ 8K 0.29 GB Runtime overhead 0.6 GB Over budget 47.0 GB past 14.4 GB

Quantisation ladder

Quant	Weights	Total @ 8K	Max context	Tok/s	Quality	Fit
MXFP4	60.5 GB	61.4 GB	—	~6.7	Reference	47.0 GB over

Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. This model interleaves sliding-window layers (128 tokens, 1 global in 2), which is why its cache barely grows with context.

How to run it

Ollama llama.cpp LM Studio vLLM

terminal

$ llama-server \
    -hf openai/gpt-oss-120b:MXFP4 \
    -c 8192 -ngl 8

The engine underneath most of the others. Every knob is exposed. More on llama.cpp.

01Download is 60.5 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.

02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.

03If it falls back to CPU silently, drop the context first, then step down a quantisation.

See all models for this rig Compare with another model