OpenAI's gpt-oss-20b fits on a 16GB graphics card and flies at 225 tokens/second on an RTX 4090. But there's a context-length trap that quietly eats your VRAM — here's what you actually need.
Yes — gpt-oss-20b runs locally on a single 16GB graphics card, and it's fast. At OpenAI's default MXFP4 quantization with a sensible 8K context, it needs only about 12.7GB of [VRAM](/how-much-vram-do-you-need-to-run-ai-models-2026) (roughly 11.2GB weights plus KV cache and overhead), so it fits comfortably on a [16GB card](/best-gpu-for-gpt-oss-20b-2026) like an RTX 4080, 4090, or 5060 Ti 16GB — and it flies at about 225 tokens/second on an [RTX 4090](/best-gpu-for-local-ai-2026). There's one catch worth knowing before you buy: the model supports up to 128K context, and pushing toward that balloons the VRAM (the KV cache grows with context), so 16GB is great for normal use but 24GB is the safer pick if you want long documents. Here's exactly what you need.
The numbers: VRAM, speed, and the context trap
Let's be precise, because the details decide your card. At OpenAI's MXFP4 quantization (its native 4-bit format) with 8K context, gpt-oss-20b uses about 12.7GB of VRAM — comfortably inside 16GB. That's the config most people should run, and on an RTX 4090 it generates around 225 tokens/second, which is very fast — far quicker than you can read. (For reference, running it at full BF16 precision would need ~42GB just for weights, which is why the 4-bit MXFP4 build is the one to use locally.) Now the catch, and it's the thing that trips people up: gpt-oss-20b can handle context up to 128K tokens, but the [KV cache](/what-is-a-context-window-explained-2026) grows with context length, so feeding it a big document can push VRAM well past 16GB. In practice: 16GB is perfect for short-to-medium context (chat, coding snippets, 8K), but if you want to throw whole documents or long histories at it, you'll want 24GB so the context doesn't crowd out the model. Match your card to how you'll actually use it.
gpt-oss-20b: what runs it
16GB GPU (4080/5060 Ti 16GB)
Your setup
Yes
Runs gpt-oss-20b?
Great at 8K context; tight at long context
24GB GPU (4090 / used 3090)
Your setup
Yes, ideal
Runs gpt-oss-20b?
Headroom for 128K context
24GB+ Mac (unified memory)
Your setup
Yes
Runs gpt-oss-20b?
Slower than a GPU but works
8-12GB GPU
Your setup
Not comfortably
Runs gpt-oss-20b?
Model won't fit with context
Your setup
Runs gpt-oss-20b?
Notes
16GB GPU (4080/5060 Ti 16GB)
Yes
Great at 8K context; tight at long context
24GB GPU (4090 / used 3090)
Yes, ideal
Headroom for 128K context
24GB+ Mac (unified memory)
Yes
Slower than a GPU but works
8-12GB GPU
Not comfortably
Model won't fit with context
gpt-oss-20b needs ~12.7GB at 8K context — but the 128K-context KV cache is what pushes you to 24GB. · Unsplash
How to actually run it
The easy path is the same as any local model: install [Ollama or LM Studio](/ollama-vs-lm-studio-which-should-you-use-2026) and pull gpt-oss-20b — they handle the MXFP4 build and the setup for you, and you'll be chatting in minutes. If you have a [16GB card](/best-gpu-for-gpt-oss-20b-2026), keep your context reasonable (8K is plenty for most work) and it'll run smoothly at high speed. If you have a [24GB card](/used-rtx-3090-buying-guide-local-ai-2026) (an RTX 4090 or a used 3090), you get the same speed plus the headroom to use long context without VRAM anxiety — which is why 24GB is the sweet spot for this model. On a Mac, gpt-oss-20b runs via unified memory (a 24GB+ machine), slower than a discrete GPU but perfectly usable. Whatever you're on, size it first in the VRAM calculator at the context length you actually need — that one step tells you instantly whether your card is comfortable or on the edge. The bottom line: gpt-oss-20b is one of the most accessible strong models you can run at home — 16GB gets you in the door, 24GB gives you room to breathe.
Quick answers
How much VRAM do you need to run gpt-oss-20b?
At OpenAI's default MXFP4 quantization with an 8K context window, gpt-oss-20b needs about 12.7GB of VRAM (roughly 11.2GB for weights plus KV cache and overhead), so it fits on a single 16GB GPU. The important caveat is context length: the model supports up to 128K tokens, but the KV cache grows as you use more context, which can push VRAM past 16GB. So 16GB is comfortable for everyday short-to-medium context use, while 24GB is the safer choice if you want to feed it long documents. Running the model at full BF16 precision would need around 42GB, which is why the 4-bit build is the one to use locally.
How fast is gpt-oss-20b locally?
Very fast on a modern GPU — about 225 tokens per second on an RTX 4090 with an 8K context window. That's far quicker than you can read, so responses feel instant. The speed comes partly from its Mixture-of-Experts design, which activates only a fraction of its parameters per token. Speed does drop as you increase context length (more tokens to process) and on less powerful cards, but even mid-range 16GB GPUs run it at very usable rates. On a Mac with unified memory it runs more slowly than a discrete GPU but is still perfectly usable for everyday tasks.
What GPU is best for gpt-oss-20b?
A 16GB card is the entry point and a 24GB card is the sweet spot. On 16GB (an RTX 4080, 4090, or the cheaper RTX 5060 Ti 16GB) the model runs great at normal context lengths, but you can run short on VRAM if you push toward its 128K context. A 24GB card — an RTX 4090 or a used RTX 3090 for value — gives you the same speed plus headroom for long context, which is why it's the recommended pick for this model. Avoid 8-12GB cards, which can't comfortably fit the model plus context. A 24GB+ Mac also works via unified memory, just more slowly.
