Only with CPU offload

gpt-oss 20B at MXFP4 needs 11.6 GB but only 10.6 GB is addressable, so about 9% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 32 tokens per second — usable for batch work, painful for chat.

91% on GPU 8K context MXFP4 · 10.8 GB Apache 2.0 Released Aug 2025 Native MXFP4

Ships natively in MXFP4, so the 4-bit weights are the reference weights, not a lossy copy. Fits 16 GB.

What hardware do I need for gpt-oss 20B? →

The VRAM budget

weights 10.8 GB

Weights 10.8 GB KV cache @ 8K 0.19 GB Runtime overhead 0.6 GB Over budget 1.0 GB past 10.6 GB

Quantisation ladder

Quant	Weights	Total @ 8K	Max context	Tok/s	Quality	Fit
MXFP4	10.8 GB	11.6 GB	—	~32	Reference	1.0 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-20b:MXFP4 \
    -c 8192 -ngl 21

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

01Download is 10.8 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