How much VRAM does gpt-oss 20B need?
gpt-oss 20B needs 11.8 GB for its weights at 4-bit, 12.9 GB with an 8K context and 13.5 GB with 32K.
- Parameters
- 20.9B
- Active per token
- 3.6B
- Longest context
- 128K
- Released
- Aug 2025
- License
- Apache 2.0
- Strongest at
- Agents
Memory at each precision
| Precision | Weights | With 8K | With 32K | With 128K |
|---|---|---|---|---|
| As released (MXFP4) | 11.3 GB | 12.4 GB | 13.0 GB | 15.2 GB |
| 4-bit | 11.8 GB | 12.9 GB | 13.5 GB | 15.7 GB |
| 3-bit | 9.5 GB | 10.6 GB | 11.2 GB | 13.4 GB |
| 2-bit | 7.3 GB | 8.4 GB | 9.0 GB | 11.2 GB |
| 1.6-bit | 4.0 GB | 5.1 GB | 5.7 GB | 7.9 GB |
One conversation, 16-bit context cache, runtime overhead included.
What runs it
| Machine | Best fit | Context | Speed | First reply |
|---|---|---|---|---|
| RTX 3060 12 GB | 3-bit | 40K | 119 tok/s | 1.4 s |
| RTX 4060 Ti 16 GB | MXFP4 | 120K | 71 tok/s | 0.8 s |
| RTX 3090 | MXFP4 | 128K | 187 tok/s | 0.5 s |
| RTX 4090 | MXFP4 | 128K | 197 tok/s | 0.2 s |
| RTX 5090 | MXFP4 | 128K | 287 tok/s | 0.2 s |
| 2× RTX 3090 | MXFP4 | 128K | 187 tok/s | 0.3 s |
| RTX PRO 6000 Blackwell | MXFP4 | 128K | 287 tok/s | 0.1 s |
| M4 Max 128 GB | MXFP4 | 128K | 94 tok/s | 2.9 s |
| M5 Max 128 GB | MXFP4 | 128K | 102 tok/s | 0.8 s |
| Ryzen AI Max+ 395 128 GB | MXFP4 | 128K | 53 tok/s | 1.7 s |
| DGX Spark 128 GB | MXFP4 | 128K | 67 tok/s | 0.6 s |
| M3 Ultra 256 GB | MXFP4 | 128K | 122 tok/s | 1.6 s |
| M5 Ultra 512 GB | MXFP4 | 128K | 150 tok/s | 0.5 s |
| H100 SXM | MXFP4 | 128K | 395 tok/s | 0.1 s |
| DGX Station GB300 | MXFP4 | 128K | 511 tok/s | 0.1 s |
| 8× H200 NVL | MXFP4 | 128K | 454 tok/s | 0.1 s |
Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.
How to run it
- Ollama
ollama run gpt-oss:20b- llama.cpp
llama-server -hf unsloth/gpt-oss-20b-GGUF- vLLM
vllm serve openai/gpt-oss-20b- MLX (Mac)
mlx_lm.chat --model mlx-community/gpt-oss-20b-MXFP4-Q4
Questions
How much VRAM does gpt-oss 20B need?
gpt-oss 20B needs 11.8 GB for its weights at 4-bit, 12.9 GB with an 8K context and 13.5 GB with 32K.
Can gpt-oss 20B run on an RTX 4090?
Yes. At MXFP4 it holds 128K of context and writes about 197 tokens a second.