Can I Run gpt-oss-20b on a 16 GB GPU?

Short answer: yes on 16 GB, no on 12 GB. OpenAI’s gpt-oss-20b is a Mixture-of-Experts model with about 20.9 billion parameters in total. All of them must be stored in memory, which comes to about 12.8 GB at 4-bit-class size, and with an 8K context the total is about 14.0 GB.

Check your own card with the Can My GPU Run This AI Model? calculator.

Why the KV cache is so small

gpt-oss-20b has 24 layers, but only 12 of them use full attention. The other 12 use a sliding window of 128 tokens. With 8 KV heads of size 64, the KV cache costs about 0.20 GB at 8K context, far less than a dense model of similar size. The memory is dominated by the weights.

Which GPUs can run it?

Because it is a Mixture-of-Experts model, only a few experts run per token (4 of 32). Real speed is usually higher than a dense model of the same size, so we do not estimate it here.

What about 12 GB cards?

At about 14 GB, gpt-oss-20b does not fit 12 GB cards. Layers or experts can be offloaded to system RAM, but speed will drop. If you are on 12 GB, Qwen3 14B or Gemma 3 12B are better fits.

Frequently asked questions

Can an RTX 4060 Ti 16GB run gpt-oss-20b?

Yes, about 14 GB at 8K context.

Can an 8 GB card run it?

Not on the GPU alone.

Is gpt-oss-20b a dense model?

No. It is a Mixture-of-Experts model with 32 experts, 4 active per token.

Sources and method

• Architecture: gpt-oss-20b config on Hugging Face (24 layers, 12 full and 12 sliding-window, 32 experts with 4 per token).
• The weights are shipped in a 4-bit format; we use our Q4 size estimate, so the real file may differ by a GB or so.

Estimates: weights = parameters × bits per weight; KV cache from the model’s published architecture; about 1 GB of runtime overhead. “Fits” means at most 92 percent of VRAM, “Tight” is 92 to 100 percent. Speed is memory bandwidth divided by model size at 60 to 80 percent efficiency. These are planning estimates, not our own benchmarks.

Last updated 3 Nov 2026.