# gpt-oss-20b

> OnPrem World · https://onpremworld.com/en/models/gpt-oss-20b · Last updated: 2026-10-14 · CC BY 4.0

OpenAI's open-weight 21B MoE with 3.6B active and adjustable reasoning effort (low, medium, high); it fits a 16 GB card.

|  |  |
|---|---|
| Developer | OpenAI |
| Parameters | 21B (3.6B active, MoE) |
| Native context | 131,072 tokens |
| Licence | Apache-2.0 |
| Commercial use | Yes |
| KV cache | 12 global + 12 sliding-window layers (128 tokens) |
| Released | 2025-08 |

```
ollama run gpt-oss:20b
```

## Memory needed

Weights + KV cache + runtime overhead, in GiB. Lower quantization = less memory, slightly lower quality.

| Quantization | 4K | 8K | 32K |
|---|---|---|---|
| Q4_K_M | 13.0 GiB | 13.1 GiB | 13.7 GiB |
| Q5_K_M | 15.2 GiB | 15.3 GiB | 15.9 GiB |
| Q6_K | 17.5 GiB | 17.6 GiB | 18.2 GiB |
| Q8_0 | 22.4 GiB | 22.5 GiB | 23.1 GiB |
| FP16 | 41.7 GiB | 41.8 GiB | 42.3 GiB |

## Which hardware runs it (8K)

| Hardware | Memory | Best quant | tok/s (est.) |
|---|---|---|---|
| GeForce RTX 5090 | 32 GB | Q8_0 | 328 |
| GeForce RTX 5080 | 16 GB | Q5_K_M | 262 |
| GeForce RTX 5070 Ti | 16 GB | Q5_K_M | 245 |
| GeForce RTX 4080 SUPER | 16 GB | Q5_K_M | 201 |
| Radeon RX 7900 XT | 20 GB | Q6_K | 189 |
| GeForce RTX 4090 | 24 GB | Q8_0 | 184 |
| Radeon RX 7900 XTX | 24 GB | Q8_0 | 176 |
| Radeon RX 9070 XT | 16 GB | Q5_K_M | 175 |
| RTX PRO 6000 Blackwell Workstation Edition | 96 GB | FP16 | 174 |
| GeForce RTX 3090 | 24 GB | Q8_0 | 171 |
| RTX PRO 5000 72GB Blackwell | 72 GB | FP16 | 131 |
| GeForce RTX 5060 Ti 16GB | 16 GB | Q5_K_M | 122 |
| Radeon AI PRO R9700 | 32 GB | Q8_0 | 117 |
| Arc Pro B70 | 32 GB | Q8_0 | 111 |
| M5 Ultra | 512 GB | FP16 | 108 |
| RTX 6000 Ada Generation | 48 GB | FP16 | 93 |
| Radeon RX 9060 XT 16GB | 16 GB | Q5_K_M | 87 |
| Arc Pro B60 | 24 GB | Q8_0 | 83 |
| GeForce RTX 4060 Ti 16GB | 16 GB | Q5_K_M | 79 |
| M5 Max (18-core CPU / 32-core GPU) | 36 GB | Q8_0 | 78 |
| RTX A6000 | 48 GB | FP16 | 75 |
| M3 Ultra | 512 GB | FP16 | 74 |
| M4 Max (14-core CPU / 32-core GPU) | 36 GB | Q8_0 | 70 |
| M5 Max (18-core CPU / 40-core GPU) | 128 GB | FP16 | 55 |
| M4 Max (16-core CPU / 40-core GPU) | 128 GB | FP16 | 49 |
| M6 (24GB/32GB configurations, 170 GB/s) | 32 GB | Q8_0 | 29 |
| M5 Pro | 64 GB | FP16 | 28 |
| M5 | 32 GB | Q8_0 | 26 |
| RTX Spark (N1X) laptops & compact desktops | 128 GB | FP16 | 25 |
| M4 Pro | 64 GB | FP16 | 25 |
| DGX Spark | 128 GB | FP16 | 23 |
| Ryzen AI Max+ PRO 495 (Gorgon Halo) systems | 192 GB | FP16 | 23 |
| Ryzen AI Max+ 395 (Strix Halo) systems | 128 GB | FP16 | 21 |
| GeForce RTX 3060 (12 GB) | 12 GB | Doesn't fit | – |
| GeForce RTX 5070 | 12 GB | Doesn't fit | – |
| Arc B580 | 12 GB | Doesn't fit | – |
| M6 (16GB configuration, 153 GB/s) | 16 GB | Doesn't fit | – |

## Frequently asked questions

### How much VRAM does gpt-oss-20b need?

With an 8K context, about 13.1 GiB at Q4_K_M and 22.5 GiB at Q8_0 (weights + KV cache + runtime overhead). Longer contexts need more memory for the KV cache.

### Can gpt-oss-20b run on an GeForce RTX 4090?

Yes. gpt-oss-20b fits on the GeForce RTX 4090 (24 GB) at Q8_0 with an 8K context, using about 22.5 GiB, at an estimated 184 tokens/s.

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

Yes. gpt-oss-20b fits on the M3 Ultra (96 GB) at FP16 with an 8K context, using about 41.8 GiB, at an estimated 74 tokens/s.

### Can gpt-oss-20b be used commercially?

gpt-oss-20b is released under Apache-2.0. The licence allows commercial use (read its conditions).

## Sources

- https://openai.com/index/introducing-gpt-oss/
- https://huggingface.co/openai/gpt-oss-20b
- https://huggingface.co/openai/gpt-oss-20b/resolve/main/config.json
- https://ollama.com/library/gpt-oss/tags