# gpt-oss-20b: hardware requirements
> gpt-oss-20b (OpenAI) is a 20.9B (3.6B active) mixture-of-experts model with a 128K-token context window, released under the Apache 2.0 license. It needs about 15GB of memory at MXFP4 with 8K context.
gpt-oss-20b is OpenAI's small open-weight reasoning model, a 21B mixture-of-experts with 3.6B active parameters whose expert weights ship natively in 4-bit MXFP4. OpenAI has not released a successor, and it remains one of the best reasoning and tool-calling models that fits a 16GB GPU. The native weights take about 14GB, so 16GB cards and Macs with 24GB or more run it fully in memory, and 12GB cards run it with a little CPU offload. Reasoning effort is adjustable (low, medium, high) but cannot be turned off.

## Recommended hardware
- **Minimum GPU:** RTX 4060 Ti 16GB (16GB), MXFP4, ~51 tokens/sec
- **Best value:** RTX 3090 (24GB), MXFP4, ~120 tokens/sec
- **Fastest single GPU:** RTX 5090 (32GB), MXFP4, ~190 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), MXFP4, ~41 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| MXFP4 | 13.8GB | 15.0GB | 15.6GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX 5090 | 32GB | MXFP4 | ~190 tok/s | ~$4,300 |
| RTX PRO 6000 | 96GB | MXFP4 | ~190 tok/s | ~$16,000 |
| RTX PRO 5000 72GB | 72GB | MXFP4 | ~160 tok/s | ~$9,100 |
| RTX 4090 | 24GB | MXFP4 | ~135 tok/s | ~$2,500 used |
| RTX 5080 | 16GB | MXFP4 | ~130 tok/s | ~$1,600 |
| RTX 6000 Ada | 48GB | MXFP4 | ~130 tok/s | ~$4,750 used |
| RTX 5070 Ti | 16GB | MXFP4 | ~125 tok/s | ~$1,200 |
| RTX 3090 | 24GB | MXFP4 | ~120 tok/s | ~$1,050 used |
| RTX 4080 Super | 16GB | MXFP4 | ~110 tok/s | ~$1,150 used |
| M5 Ultra 512GB | 512GB | MXFP4 | ~105 tok/s | n/a |
| M5 Ultra 256GB | 256GB | MXFP4 | ~105 tok/s | ~$9,499 |
| RTX A6000 | 48GB | MXFP4 | ~105 tok/s | ~$4,400 used |
| V100 32GB | 32GB | MXFP4 | ~100 tok/s | ~$700 used |
| RTX 4070 Ti Super | 16GB | MXFP4 | ~100 tok/s | ~$900 used |
| RX 7900 XTX | 24GB | MXFP4 | ~99 tok/s | ~$1,050 used |
| M3 Ultra 512GB | 512GB | MXFP4 | ~80 tok/s | $9,499 MSRP |
| R9700 | 32GB | MXFP4 | ~75 tok/s | ~$1,800 |
| RX 9070 XT | 16GB | MXFP4 | ~75 tok/s | ~$800 |
| RTX 5060 Ti 16GB | 16GB | MXFP4 | ~74 tok/s | ~$620 |
| MI50 32GB | 32GB | MXFP4 | ~74 tok/s | ~$600 used |
| M5 Max 128GB | 128GB | MXFP4 | ~73 tok/s | ~$5,100 |
| M4 Max 128GB | 128GB | MXFP4 | ~67 tok/s | $3,499 MSRP |
| RTX 4000 Ada | 20GB | MXFP4 | ~62 tok/s | $1,250 MSRP |
| Arc Pro B70 | 32GB | MXFP4 | ~59 tok/s | ~$1,300 |
| RTX 4060 Ti 16GB | 16GB | MXFP4 | ~51 tok/s | ~$380 used |

## How to run it
- Ollama: `ollama run gpt-oss:20b`
- llama.cpp: `llama-server -hf ggml-org/gpt-oss-20b-GGUF -c 8192 -ngl 99`
- MLX (Mac): `mlx_lm.chat --model mlx-community/gpt-oss-20b-MXFP4-Q8`
- Weights: https://huggingface.co/openai/gpt-oss-20b

## Strengths
- Strong math, reasoning, and tool calling for a model that fits 16GB
- Ships in native MXFP4, so there is no quantization quality loss to manage
- Very fast: 3.6B active parameters and a tiny KV cache (24 KiB per token)
- Apache 2.0 license

## Weaknesses
- Thin world knowledge and more hallucination than dense models of similar size
- Reasoning cannot be disabled, which adds latency to simple requests
- Text only, and it relies on the harmony chat format, which some tools mishandle
Full page: https://llamabuilds.ai/models/gpt-oss-20b · Hardware guide: https://llamabuilds.ai/models/gpt-oss-20b/requirements

---
Source: https://llamabuilds.ai/models/gpt-oss-20b · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.