OpenAI · gpt-oss · 20.9B parameters · 3.6B active
gpt-oss 20B-A3.6B VRAM requirements
gpt-oss 20B-A3.6B has 24 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 11.9 GB, and the best quantisation that fits a 24 GB card is INT8 / W8A8.
Runs comfortably
13.1 GB of 21.8 GB · 60%gpt-oss 20B-A3.6B at Q4_K_M leaves 8.7 GB spare on a RTX 4090. There is room to raise the context length or move up a quantisation level.
Every quantisation of gpt-oss 20B-A3.6B on a RTX 4090
Highlighted row is the highest quality that still fits at 8K context.
| Quantisation | bpw | Weights | Total | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|---|
| FP16 / BF16 | 16.00 | 38.9 GB | 40.1 GB | Won't fit | — | 10.2 |
| INT8 / W8A8 | 8.50 | 20.4 GB | 21.6 GB | Fits, but tight | 12K | 123 |
| FP8 (E4M3) | 8.00 | 19.5 GB | 20.6 GB | Fits, but tight | 32K | 128 |
| Q6_K | 6.56 | 16.0 GB | 17.1 GB | Runs comfortably | 107K | 150 |
| Q5_K_M | 5.67 | 13.8 GB | 15.0 GB | Runs comfortably | 128K | 168 |
| Q5_K_S | 5.52 | 13.4 GB | 14.6 GB | Runs comfortably | 128K | 172 |
| AWQ 4-bit | 4.25 | 11.9 GB | 13.1 GB | Runs comfortably | 128K | 188 |
| GPTQ 4-bit | 4.25 | 11.9 GB | 13.1 GB | Runs comfortably | 128K | 188 |
| MXFP4 | 4.25 | 11.9 GB | 13.1 GB | Runs comfortably | 128K | 188 |
| Q4_K_M | 4.85 | 11.9 GB | 13.1 GB | Runs comfortably | 128K | 188 |
| Q4_K_S | 4.58 | 11.3 GB | 12.4 GB | Runs comfortably | 128K | 196 |
| Q4_0 | 4.55 | 11.2 GB | 12.4 GB | Runs comfortably | 128K | 196 |
| IQ4_XS | 4.25 | 10.5 GB | 11.7 GB | Runs comfortably | 128K | 206 |
| Q3_K_M | 3.91 | 9.7 GB | 10.9 GB | Runs comfortably | 128K | 217 |
| IQ3_M | 3.70 | 9.2 GB | 10.4 GB | Runs comfortably | 128K | 225 |
| IQ3_XXS | 3.06 | 7.8 GB | 8.9 GB | Runs comfortably | 128K | 253 |
| Q2_K | 2.63 | 6.8 GB | 8.0 GB | Runs comfortably | 128K | 276 |
| IQ2_XXS | 2.06 | 5.5 GB | 6.7 GB | Runs comfortably | 128K | 313 |
| IQ1_M | 1.75 | 4.8 GB | 5.9 GB | Runs comfortably | 128K | 338 |
gpt-oss 20B-A3.6B on each GPU
Q4_K_M weights at 8K context, single card, monitor attached.
| GPU | VRAM | GB/s | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|
| H100 SXM 80GB | 80 | 3350 | Runs comfortably | 128K | 470 |
| A100 80GB | 80 | 2039 | Runs comfortably | 128K | 327 |
| RTX 5090 | 32 | 1792 | Runs comfortably | 128K | 311 |
| RTX 5080 | 16 | 960 | Fits, but tight | 33K | 195 |
| RTX 4090 | 24 | 1008 | Runs comfortably | 128K | 188 |
| RTX 5070 Ti | 16 | 896 | Fits, but tight | 33K | 185 |
| RTX 3090 | 24 | 936 | Runs comfortably | 128K | 183 |
| Radeon RX 7900 XTX | 24 | 960 | Runs comfortably | 128K | 173 |
| L40S | 48 | 864 | Runs comfortably | 128K | 166 |
| RTX A6000 | 48 | 768 | Runs comfortably | 128K | 156 |
| Mac Studio M3 Ultra 256GB | 256 | 819 | Runs comfortably | 128K | 154 |
| RTX 4080 Super | 16 | 736 | Fits, but tight | 33K | 145 |
| RTX 4070 Ti Super | 16 | 672 | Fits, but tight | 33K | 134 |
| Mac Studio M4 Max 128GB | 128 | 546 | Runs comfortably | 128K | 119 |
| RTX 5060 Ti 16GB | 16 | 448 | Fits, but tight | 33K | 102 |
| NVIDIA DGX Spark (GB10) | 128 | 273 | Runs comfortably | 128K | 64.8 |
| Mac Mini M4 Pro 48GB | 48 | 273 | Runs comfortably | 128K | 63.2 |
| RTX 4060 Ti 16GB | 16 | 288 | Fits, but tight | 33K | 62.2 |
| RTX 5070 | 12 | 672 | Won't fit | — | 53.6 |
| Ryzen AI Max+ 395 128GB | 128 | 256 | Runs comfortably | 128K | 49.1 |
| RTX 4070 Super | 12 | 504 | Won't fit | — | 45.3 |
| RTX 4070 | 12 | 504 | Won't fit | — | 45.3 |
| RTX 3060 12GB | 12 | 360 | Won't fit | — | 42.3 |
| Arc B580 | 12 | 456 | Won't fit | — | 37.5 |
| RTX 3080 10GB | 10 | 760 | Won't fit | — | 36.3 |
Architecture
| Parameters | 20.9B |
| Active per token | 3.6B of 32 experts, top-4 |
| Layers | 24 |
| Hidden size | 2880 |
| Attention heads / KV heads | 64 / 8 |
| Head dimension | 64 |
| Vocabulary | 201,088 |
| Trained context | 128K |
| KV cache per 1K tokens | 0 GB |
| Hugging Face | openai/gpt-oss-20b |
The OpenAI family
gpt-oss 20B and 120B are mixture-of-experts models that ship natively in MXFP4, so the quantisation ladder starts at the published 4-bit weights rather than at fp16. Both read very few parameters per token — 3.6B and 5.1B — which makes them unusually fast for their size.
huggingface.co/openai · github.com/openai/gpt-oss · all 2 OpenAI models
Direct answers
See the verdictgpt-oss 20B-A3.6B on RTX 4090
See the verdictgpt-oss 20B-A3.6B on RTX 3090
See the verdictgpt-oss 20B-A3.6B on RTX 5080
See the verdictgpt-oss 20B-A3.6B on RTX 5070 Ti
See the verdictgpt-oss 20B-A3.6B on RTX 5070
See the verdictgpt-oss 20B-A3.6B on RTX 4070 Ti Super
See the verdictgpt-oss 20B-A3.6B on RTX 4070 Super
See the verdictgpt-oss 20B-A3.6B on RTX 4070
See the verdict