Which GPU runs which model · OpenAI gpt-oss
Which GPU runs gpt-oss-20b, and what it costs per hour
gpt-oss-20b (21B MoE (3.6B active)) takes about 42 GB of VRAM as FP16 weights and about 13 GB at native MXFP4, before the KV cache for your context window. The table lists the cards on cloudgpu.app by VRAM class, whether the model fits, and the price you will actually be billed, read on 2026-09-15.
Recommended card: RTX 3090 24G (24 GB) at $0.21 per hour, the lowest-priced card in a VRAM class that fits.
| VRAM | Cards | Price per hour | Fit |
|---|---|---|---|
| 24 GB | RTX 3090 24G · RTX 4090D 24G · RTX 4090 24G | $0.21 – $0.33/h | Fits MXFP4 fits comfortably |
| 32 GB | RTX 5090 32G | $0.59/h | Fits MXFP4 fits |
| 48 GB | RTX A6000 48G · RTX 4090 48G | $0.44 – $0.65/h | Fits fits |
| 80 GB | A800 80G · A100 80G | $1.09 – $1.49/h | Fits fits with batching |
| 96 GB | H20 96G · RTX PRO 6000 96G | $1.29 – $1.49/h | Fits fits with batching |
Prices are the retail rates on cloudgpu.app read 2026-09-15 and equal what the billing system charges; VRAM figures were checked on 2026-09-08. Stock changes during the day.
Reading the table
"Fits" is judged for the native MXFP4 weights plus a working context window; FP16 needs roughly twice the VRAM. The note under each verdict says how much context you get.
Data-centre cards (80 GB and 96 GB) are not cheap on cloudgpu.app. If the model fits on a consumer card, the consumer card is the better deal. The prices here are the same numbers the billing system charges.
Questions about running gpt-oss-20b
How much VRAM does gpt-oss-20b need?
gpt-oss-20b has 21B MoE (3.6B active) parameters. The FP16 weights take about 42 GB; at native MXFP4 the weights take about 13 GB. Add 1 to 8 GB for the KV cache depending on context length and batch size.
What is the cheapest GPU that runs gpt-oss-20b?
On cloudgpu.app the lowest-priced card we recommend for gpt-oss-20b is the RTX 3090 24G with 24 GB at $0.21 per hour (MXFP4 fits comfortably). Prices read 2026-09-15; the pricing page shows live stock.
Can I run gpt-oss-20b on an RTX 4090 with 24 GB?
Yes. On a 24 GB card: MXFP4 fits comfortably. An RTX 4090 24 GB is $0.33 per hour on cloudgpu.app; the RTX 3090 24 GB at $0.21 per hour has the same VRAM and is slower but cheaper.
What does it cost per hour to run gpt-oss-20b on cloudgpu.app?
From $0.21 per hour on the RTX 3090 24G, billed per minute. A 20-minute test is about $0.07; one hour is pre-authorised when the instance starts and unused minutes are refunded on stop. Cards in higher VRAM classes cost more; the table above lists each.
How do I start gpt-oss-20b on a rented GPU?
Open Rent a GPU, choose the Ollama template and a card from the table above. The instance is ready in about 100 seconds with a public HTTPS endpoint; then run "ollama pull gpt-oss:20b" in the terminal, or request model "gpt-oss:20b" through the API and Ollama downloads it on first use.
Do I need a credit card or a Chinese account?
No Chinese account, phone number or ID is required. You can pay by card (Visa or Mastercard, any amount from $5) or USDT (TRC-20). A first top-up of $10 or more gets a $5 bonus, so a $10 card payment starts you with $15 of GPU time.