Best GPU for Running gpt-oss Locally in 2026 (20B and 120B)

People ask me for 'the best GPU for gpt-oss' as if it's one question, but OpenAI shipped two very different models under that name, and they land on opposite ends of the hardware world. The 20B is a genuine joy — it fits on a normal gaming card and runs fast. The 120B is a different conversation entirely: it's an 80GB-class model that most people will rent or offload rather than own. The reason both are friendlier than their parameter counts suggest is that OpenAI released them in MXFP4 (a native 4-bit format), so they're far smaller in memory than a typical FP16 model of the same size. So let me split this cleanly — here's the card I'd buy for the 20B, and the honest reality of the 120B.

gpt-oss-20B: a great fit for a normal card

This is the one most people actually want, and it's refreshingly undemanding. Because it ships in MXFP4, gpt-oss-20B loads in about 14GB rather than the ~40GB an FP16 20B would need. In practice that makes 16GB my hard floor — enough to hold the weights with a bit of room — while 24GB gives you comfortable context headroom. Speed is not the problem here: reviewers clock it around 225 tokens/sec on an RTX 4090, which is faster than you can read. On a 12GB card it still runs, but part of the model offloads to system RAM and the speed drops off a cliff — I wouldn't build around that. Watch the 128k context window, too: long contexts eat extra memory on top of the weights, so that's where a 24GB card earns its keep.

gpt-oss-120B: the 80GB problem

Here's where the mixture-of-experts design trips people up. gpt-oss-120B has 116.8B parameters but only activates ~5.13B per token — and I keep having to tell people that the active number does not size your hardware. Every expert weight has to sit in memory even if only a few fire per token, so you plan for the whole thing. OpenAI says the ~60.8GiB MXFP4 checkpoint can run on a single 80GB GPU, and the roughly-60GB figure is a constrained floor, not comfortable headroom. Translation: this is an 80GB-class model. On the consumer side that means something like an RTX PRO 6000 Blackwell (96GB) — glorious, but priced like a car — or you split it across multiple cards, which adds real complexity. My honest answer for almost everyone: don't own it.

Which GPU I'd actually buy

For the 20B — which is what most of you are really after — I'd buy a 16GB card and not overthink it. On pure value my pick is a used RTX 3090 (24GB), which gives you both the 20B with full context headroom and room to grow; the used-GPU value case is here. If you want new, an RX 9070 or RTX 5070 Ti (both 16GB) is plenty. For the 120B, I would not buy a card at all unless you run it daily: rent an 80GB H100 by the hour (I compared the neocloud rental prices), or go the unified-memory route for a slower-but-owned option. And whatever you're eyeing, size it against your real target first in the cost-to-run tool and the wider best-GPU-for-local-AI pillar — the deeper 120B hardware breakdown lives in the gpt-oss-120B requirements guide.