How much VRAM gpt-oss 20B (MoE) needs, and what it costs to run
Some links on this page are referral links: we may earn a commission, at no cost to you.
Trying to decide what to buy? The best GPU for gpt-oss 20B (MoE) ↓ · Got a card already? See what your PC can run →
Cheapest way to rent it right now
$0.03/hr
NVIDIA L4 · MXFP4 · spot · ≈ $8/mo at 8h/day
Live price captured Thu, 19 Nov 2026 18:20:07 GMT.It never changes which card is cheapest.
Memory needed, by quantisation
At 16k context. Weights + KV cache + ~0.8 GB overhead. gpt-oss 20B (MoE) ships in MXFP4, so its own release leads and no bigger copy is listed: it would only store the same numbers in more bytes. Memory is in binary GB (1,024³ bytes), the unit GPU memory is sold in; Hugging Face and Ollama list downloads in decimal GB, about 7% more.
| Quant | Weights | KV cache | Total | Quality |
|---|---|---|---|---|
| MXFP4official | 12.8 GB | 0.8 GB | 14.3 GB | The model's own release (MXFP4, a 13.8 GB download): the best quality there is. A bigger copy only stores the same numbers in more bytes. Its 4-bit GGUF builds keep the 4-bit experts as they are, so they come out about this size too. |
Cards that can run it
Best quantisation each card fits at 16k context, with live rental price where we track one.
The best GPU for gpt-oss 20B (MoE)
Ranked from live rental prices and computed VRAM fit, not opinion. Every card is judged at its own MXFP4 release with a 8K context, so the comparison is fair; a card that could only run a crushed version isn't listed.
Best value
NVIDIA GeForce RTX 5060 Ti 16GB
Best throughput per dollar among cards that fit.
Cheapest that runs it
NVIDIA GeForce RTX 5060 Ti 16GB
Lowest hourly rental that fits it — $0.042/hr, MXFP4.
No compromise
—
Most headroom and highest throughput.
Every card that runs it, ranked
Sorted by tokens/sec per rental dollar. Throughput is a bandwidth-derived estimate.
| GPU | VRAM | VRAM used | ~tok/s | Cheapest rental | tok/s per $ |
|---|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | 14 GB15% | — | $0.400 | — |
| NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition | 96 GB | 14 GB15% | — | $0.333 | — |
| NVIDIA RTX 6000 Ada Generation | 48 GB | 14 GB29% | — | $0.269 | — |
| NVIDIA RTX PRO 5000 Blackwell | 48 GB | 14 GB29% | — | $0.163 | — |
| NVIDIA RTX A6000 | 48 GB | 14 GB29% | — | $0.200 | — |
| AMD Radeon PRO W7900 | 48 GB | 14 GB29% | — | — | — |
| NVIDIA RTX PRO 4500 Blackwell | 32 GB | 14 GB44% | — | $0.200 | — |
| NVIDIA GeForce RTX 5090 | 32 GB | 14 GB44% | — | $0.200 | — |
| NVIDIA RTX 5000 Ada Generation | 32 GB | 14 GB44% | — | $0.265 | — |
| NVIDIA GeForce RTX 3090 Ti | 24 GB | 14 GB58% | — | $0.173 | — |
| AMD Radeon RX 7900 XTX | 24 GB | 14 GB58% | — | — | — |
| NVIDIA RTX PRO 4000 Blackwell | 24 GB | 14 GB58% | — | $0.100 | — |
The value column ranks renting, because that's what we can price precisely (prices captured Thu, 19 Nov 2026 18:20:07 UTC). Buying a card? The fit and throughput columns are what you need; we don't publish a purchase-price ranking because we don't hold verified street prices. Tokens/sec is an estimate, not a benchmark. Got a card already? See everything your PC can run.
gpt-oss-20bHow these numbers are produced
- Model shape comes from the model's own config.json — 24 layers, 64 attention heads, 8 KV heads.
- Weight size is parameters × effective bits-per-weight. Those constants are checked against real published quantised file sizes — median error 0.7% across 40 measurements.
- KV cache is 2 × layers × KV-heads × head-dim × context × 2 bytes. Using KV-heads rather than attention heads is what makes this correct for grouped-query attention; treating a GQA model as multi-head overstates the cache by up to 8×.
- Tokens/sec is an estimate, not a benchmark. Generation is memory-bandwidth-bound, so this is bandwidth ÷ weight-bytes derated to 75%. Real throughput depends on your runtime, batch size and kernels. We don't run our own hardware tests — see our editorial policy.
- This is a mixture-of-experts model. Generation reads only the experts active for each token (4 of 32), so throughput is scored on ~3.6B active parameters, not the full 21B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.Active = 3.6B, vendor figure: OpenAI model card (huggingface.co/openai/gpt-oss-20b, read 29 Oct 2026): "21B parameters with 3.6B active parameters". params = HF safetensors total. Also set in workers/prices/src/models.ts seeds so the daily ingest keeps it.
- Some memory-bandwidth figures are not yet independently verified.Where that's the case we show no tokens/sec at all rather than a number we can't stand behind.
- Rental prices are pulled hourly from provider APIs (last updated Thu, 19 Nov 2026 18:20:07 GMT). Spot/interruptible pricing can change or vanish without notice.
Try it interactively
Drag the context slider and watch the KV cache fill the card
Running gpt-oss 20B (MoE): common questions
How much VRAM do you need to run gpt-oss 20B (MoE)?
Can an RTX 4090 run gpt-oss 20B (MoE)?
Can an RTX 3090 run gpt-oss 20B (MoE)?
Can an RTX 4060 Ti 16GB run gpt-oss 20B (MoE)?
What's the cheapest way to run gpt-oss 20B (MoE)?
Worth reading before you buy
Can gpt-oss-20b run on a 16GB Mac mini? Yes, but get 24GB
llama.cpp's own guide runs it on a 16GB Mac with one flag. But on a Mac the CPU and GPU share one pool of memory, so the whole model still has to fit, and 16GB leaves almost nothing for macOS. Here's the maths.
Can gpt-oss-20b run on an 8GB or 12GB GPU? Yes, with --n-cpu-moe
Everyone says 16GB minimum. That's only true if the whole model has to live on the GPU. llama.cpp's own guide runs it on an 8GB RTX 2060. Here are the commands, and the starting values for your card.
Cheapest RTX 4090 rental
Live per-hour 4090 prices on Vast and Runpod, the spot-vs-on-demand lever, and what a 24 GB card actually runs — the rent-a-card answer for everything in the 14B–32B class.
What --n-cpu-moe actually does
The trick that makes MoE models 5× faster — except it usually makes them slower, and the famous speedup only happens when the model didn't fit in the first place.