Can you run gpt-oss-20b on your GPU? Yes with 16 GB of VRAM, and slower on 8 GB

Yes, you can run gpt-oss-20b on your GPU if it has 16 GB of VRAM. OpenAI says the model runs "within 16GB of memory", and the llama.cpp guide measures 14.9 GB at an 8K-token context and 15.5 GB at 32K, so any 16 GB card holds the whole model. With an 8 GB or 12 GB card it still runs, more slowly, by keeping some expert layers in system RAM. On a Mac you need at least 16 GB of unified memory. Checked November 3, 2026 on OpenAI's model card, the llama.cpp guide, Ollama and LM Studio.

Can my GPU run gpt-oss-20b?

Find your VRAM in the table. You can see it in Windows Task Manager under Performance, GPU, "Dedicated GPU memory", or in nvidia-smi on an NVIDIA card. Memory figures are from the llama.cpp guide's minimum-requirements table; the guide's maintainer notes they "could vary a little bit" with settings.

Why does a 21-billion-parameter model fit in 16 GB? OpenAI trained its mixture-of-experts weights in MXFP4, a 4-bit format (see what quantization is), and only 3.6B of the 21B parameters work on each token. The whole model still has to sit in memory, though: the guide counts 12.0 GB of model data plus 2.7 GB of compute buffers before any context.

To check a different card or context length, put it in our VRAM calculator.

How much VRAM does gpt-oss-20b need at long context?

15.5 GB at 32K tokens and 17.9 GB at the full 128K. The model itself stays the same size; what grows is the KV cache, the memory that holds your conversation (our context window explainer shows how). The guide puts it at about 0.2 GB per 8,192 tokens for this model, which is small, so the jump from 8K to 128K is only about 3 GB.

That's why the guide's own 16 GB example (an NVIDIA V100) caps context at 32K: the card is "just at the edge to fit the full context". In practice:

• Chat, coding help, short documents: 16 GB is plenty.
• Whole books, long agent sessions, big codebases in one prompt: 24 GB, so the full 128K fits without offloading.

How fast is gpt-oss-20b on my GPU?

These are text-generation speeds (llama-bench tg128) from the benchmark runs posted in the llama.cpp guide, with the whole model on the GPU. They were run in 2025 on the guide's own builds, so treat them as a ranking, not a promise for your PC.

Even the slowest entry writes faster than most people read. If you're choosing a card to buy for it, our best GPU for gpt-oss guide puts NVIDIA's list prices next to these speeds.

Can gpt-oss-20b run on an 8 GB or 12 GB GPU?

Yes, with some of the model in system RAM. llama.cpp's --n-cpu-moe setting moves expert layers to the CPU while attention and the context stay on the GPU. The guide's example is an RTX 2060 with 8 GB: 16 MoE layers on the CPU for a 32K context, or 22 for the full context.

How slow is it? One reader posted results in the guide's comments on November 10, 2025: an RTX 4060 Laptop GPU with 8 GB, a Core i7-14700HX and 64 GB of DDR5, running 14 MoE layers on the CPU, got about 42 tokens a second. That's their machine and their settings, not ours, but it shows 8 GB is usable. A 12 GB card needs fewer layers on the CPU, so it lands between that and the full-GPU numbers above.

What else you need for offloading:

• System RAM to hold the layers the GPU doesn't. LM Studio lists 12 GB as the minimum system memory for this model; the 42 tokens/sec laptop above had 64 GB.
• The right number for --n-cpu-moe. Too low and the GPU runs out of VRAM; too high and it gets slower than it needs to be. What --n-cpu-moe actually does explains how to find it, and the step-by-step 8 GB and 12 GB setup is in gpt-oss-20b on an 8 GB or 12 GB GPU.

One Windows trap the guide warns about: the driver can let a program allocate more VRAM than the card has, then quietly swap to RAM and crawl. "Just because the model loads without errors, it doesn't mean you have enough VRAM for the settings that you are using", the guide says. Watch "GPU Memory" in Task Manager and keep it under your card's VRAM.

Can I run gpt-oss-20b on a Mac?

Yes, with 16 GB of unified memory or more. The guide says the model "can run efficiently on Macs with at least 16GB RAM", with two limits. On a 16 GB Mac, "Macs don't allow to utilize the full 16GB memory by the GPU", so you keep some layers on the CPU (the guide's example uses --n-cpu-moe 12 with a 32K context). On an 8 GB Mac, in the guide's words, "you are out of luck". Bigger Macs are quick: about 92 tokens a second on an M4 Max with 36 GB. The 16 GB Mac mini setup is covered in gpt-oss-20b on a 16 GB Mac mini, and Mac buying advice in best Mac for local AI.

How do I run gpt-oss-20b?

The easiest way is one of the two apps OpenAI's model card points to (compare them in Ollama vs LM Studio):

• Ollama: ollama pull gpt-oss:20b, then ollama run gpt-oss:20b. The download is 14 GB.
• LM Studio: lms get openai/gpt-oss-20b, or search "gpt-oss-20b" in the app.
• llama.cpp: llama-server -hf ggml-org/gpt-oss-20b-GGUF --ctx-size 0 --jinja -ub 2048 -b 2048 downloads the model and starts a local chat page and API. --ctx-size 0 means the full context, so on a 16 GB card set --ctx-size 32768 as the guide does.

The Ollama setting that cuts your context

Ollama picks a default context length from your VRAM, and on any card under 24 GB that default is only 4K tokens. A 16 GB card can handle about 32K, so raise it: use the context slider in the Ollama app's settings, or start the server with OLLAMA_CONTEXT_LENGTH=32768 ollama serve. Ollama's docs say coding tools and agents should get "at least 64000 tokens", which on gpt-oss-20b means 24 GB to stay fully on the GPU. Check what you got with ollama ps: the PROCESSOR column should say 100% GPU.

Should I rent a GPU for gpt-oss-20b instead?

Rent if you just want to try it or you only have 8 GB; buy if you'll use it every day. At 20:20 UTC on November 3, 2026, our price tracker showed single-GPU on-demand rentals on Vast from about $0.11/hr for an RTX 5060 Ti 16 GB, $0.10/hr for an RTX 3090 and $0.26/hr for an RTX 4090; Runpod listed the RTX 4090 at $0.89/hr. These are the cheapest offers at that moment and they change hourly, so check today's cloud GPU prices or the per-model page at /cost-to-run/gpt-oss-20b before you rent.

What this page can't tell you

• Your exact speed. The speeds come from benchmark runs posted in the llama.cpp guide in 2025 and a reader's laptop result, not from our own tests. Ollama, LM Studio, newer llama.cpp builds and longer contexts all change the number.
• Every card. The guide covers mostly NVIDIA and Apple hardware. AMD cards run gpt-oss through ROCm or Vulkan; the guide says "Depending on your specific hardware, the results can vary", but the memory numbers are the same.
• The bigger model. gpt-oss-120b needs about 64 GB; see gpt-oss-120b hardware requirements. To weigh gpt-oss against other open models, see gpt-oss vs Qwen3 vs DeepSeek.