Best GPU for gpt-oss-20b: VRAM Requirements
gpt-oss-20b needs about 46.4 GB of GPU memory in BF16 and 11.9 GB at 4 bits. See which GPUs fit it for inference and fine-tuning.
gpt-oss-20b needs about 46.4 GB of VRAM in FP16, 23.4 GB in FP8 and 11.9 GB at 4 bits for inference with an 8,192-token context. In FP16 it fits on one NVIDIA A100 PCIe 80GB; a 4-bit build fits on one NVIDIA GeForce RTX 5080 16GB. Its weights are published in 4-bit MXFP4, so the 4-bit figure is the closest to the checkpoint.
gpt-oss-20b is a mixture-of-experts language model with 20.9B total and 3.6B active parameters per token from OpenAI, released in August 2025, with a 131,072-token context window and published under the Apache 2.0. Its weights alone take 41.8 GB in BF16 and 10.5 GB at 4 bits per parameter; the tables below add the KV cache and runtime memory, then list the GPUs that hold each setup.
Recommended GPUs for gpt-oss-20b
Best for inference
1 GPU × NVIDIA L4 24GB
4-bit weights, 11.9 GB at 8,192 tokens.
Best for fine-tuning
1 GPU × NVIDIA A100 PCIe 80GB
LoRA on BF16 weights, 54.3 GB.
Budget pick
1 GPU × NVIDIA GeForce RTX 5080 16GB
Smallest card class that holds a 4-bit build, 11.9 GB.
gpt-oss-20b VRAM requirements by precision
One request with an 8,192-token context, KV cache in BF16, plus a 10 % margin for activations and runtime buffers.
| Precision | Weights | KV cache | Total for inference |
|---|---|---|---|
| BF16 / FP16 | 41.8 GB | 0.4 GB | 46.4 GB |
| FP8 | 20.9 GB | 0.4 GB | 23.4 GB |
| INT4 (closest to published MXFP4) | 10.5 GB | 0.4 GB | 11.9 GB |
gpt-oss-20b VRAM by task
Fine-tuning rows assume a short sequence, batch 1 and gradient checkpointing; longer sequences add activation memory.
| Task | GPU memory |
|---|---|
| Inference at the full 131,072-token context (FP8) | 30.1 GB |
| QLoRA (4-bit base) | 13.6 GB |
| LoRA (BF16 base) | 54.3 GB |
| Full fine-tuning (BF16, AdamW) | 367.8 GB |
GPUs that fit gpt-oss-20b
Number of cards whose combined memory holds each setup, assuming the model is split across them with tensor or pipeline parallelism. A dash means more than eight cards.
| GPU | Memory | BF16 inference (46.4 GB) | FP8 inference (23.4 GB) | INT4 inference (11.9 GB) | QLoRA (13.6 GB) | LoRA (54.3 GB) |
|---|---|---|---|---|---|---|
| NVIDIA GeForce RTX 5080 16GB | 16 GB | 4 GPUs | 2 GPUs | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA GeForce RTX 4090 24GB | 24 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA L4 24GB | 24 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 4 GPUs |
| NVIDIA GeForce RTX 5090 32GB | 32 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA Tesla V100 32GB | 32 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA A100 PCIe 40GB | 40 GB | 2 GPUs | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA L40S 48GB | 48 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA Quadro RTX 8000 48GB | 48 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA RTX 6000 Ada 48GB | 48 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA RTX A6000 48GB | 48 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 2 GPUs |
| NVIDIA A100 PCIe 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA A100 SXM 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA H100 PCIe 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA H100 SXM 80GB | 80 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA RTX PRO 6000 Blackwell 96GB | 96 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA H200 141GB | 141 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA B200 180GB | 180 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA B300 288GB | 288 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
| NVIDIA GB300 NVL | 288 GB | 1 GPU | 1 GPU | 1 GPU | 1 GPU | 1 GPU |
What is gpt-oss-20b?
gpt-oss-20b is the smaller of OpenAI's open-weight reasoning models: a mixture-of-experts model with 3.6 billion active parameters per token. Its expert weights ship in the 4-bit MXFP4 format, and OpenAI states that it runs within 16 GB of memory.
- Developer
- OpenAI
- Released
- August 2025
- Parameters
- 20.9B total, 3.6B active per token
- Context window
- 131,072 tokens
- License
- Apache 2.0
Expert weights are MXFP4 (about 4.25 bits per parameter) and attention weights BF16, so the 4-bit row is the closest to the published checkpoint, which takes about 13.8 GB, more than the 4-bit weight estimate. Half of the layers use a 128-token sliding window, which keeps the real KV cache below this estimate.
Source: gpt-oss-20b model card (as of November 17, 2026).
Frequently asked questions
- How much VRAM does gpt-oss-20b need?
- For inference with an 8,192-token context and one request, gpt-oss-20b needs about 46.4 GB in BF16, 23.4 GB in FP8, 11.9 GB in INT4, including a 10 % runtime margin. Real usage depends on the inference engine and its settings.
- Can gpt-oss-20b run on a single GPU?
- Yes. In BF16 it needs about 46.4 GB, which fits one NVIDIA A100 PCIe 80GB (80 GB).
- How much GPU memory does fine-tuning gpt-oss-20b take?
- QLoRA (4-bit base): about 13.6 GB; LoRA (BF16 base): about 54.3 GB; Full fine-tuning (BF16, AdamW): about 367.8 GB. Longer training sequences and larger batches add activation memory on top.
- What is the best GPU for gpt-oss-20b?
- Best for inference: 1 GPU × NVIDIA L4 24GB (4-bit weights, 11.9 GB at 8,192 tokens). Best for fine-tuning: 1 GPU × NVIDIA A100 PCIe 80GB (LoRA on BF16 weights, 54.3 GB). Budget pick: 1 GPU × NVIDIA GeForce RTX 5080 16GB (Smallest card class that holds a 4-bit build, 11.9 GB). Check live offers before booking, as availability changes.
- Can gpt-oss-20b run on an RTX 4090?
- Yes. In 4-bit it needs about 11.9 GB, which fits one RTX 4090 (24 GB). FP16 needs about 46.4 GB, 2 of these cards.
- Can gpt-oss-20b run on an RTX 5090?
- Yes. In FP8 it needs about 23.4 GB, which fits one RTX 5090 (32 GB). FP16 needs about 46.4 GB, 2 of these cards.
Estimates use the formulas of the LLM VRAM calculator (decimal GB). Real usage depends on the inference engine, quantization format and memory fragmentation: test on the target GPU before a long rental.
Explore the live market
Every offer on the live market comes from a reviewed supplier and is sold by EverXCloud Inc.