← All models
OpenAI · mixture of experts · Apache-2.0
What runs gpt-oss-20b
OpenAI's 21B MoE with 3.6B active parameters, shipped natively in MXFP4 so it fits in about 16 GB. Runs on 16 GB GPUs and most recent Macs.
- Parameters
- 21B, 3.6B active
- 4-bit file
- 10.8 GB
- MXFP4 (native)
- Needs at 8k context
- 12.5 GB
- context memory estimated
- Longest context
- 131,072 tokens
Context memory: 24,576 bytes per token for the full-attention layers only. Alternates full attention with 128-token sliding-window layers: 12 of 24 layers grow with context (2 x 12 x 8 KV heads x 64 x 2 = 24,576 bytes/token) while the other 12 hold at most 128 tokens. Card gives no per-token figure. How context memory works.
Measured on this hardware
Machines that hold the 4-bit file with 8,192 tokens of context and have published runs on the same chip or cards, fastest first. Median of single-user runs with plain decoding.
| Machine | Fit | Price | Generation |
|---|---|---|---|
| 1x RTX 5090Single GPU | Fits | ~$3,234 + unpriced parts | 290 tok/smedian of 4 runs |
| RTX 5090 + EPYC 9355P with 768 GB DDR5 (MoE offload)Hybrid GPU + RAM | Fits | ~$3,763 + unpriced parts | 290 tok/smedian of 4 runs |
| Dell Pro Max Tower T2 with NVIDIA RTX PRO 6000 BlackwellNVIDIA workstation | Fits | ~$13,909 | 287 tok/smedian of 3 runs |
| 1x RTX PRO 6000 BlackwellSingle GPU | Fits | ~$20,833 | 287 tok/smedian of 3 runs |
| Apple Mac Studio (M5 Ultra, 64-core GPU)Apple unified memory | Fitsassumed limitfrom 96 GB | ~$5,499–$9,499 | 171 tok/smedian of 1 run |
| 1x RTX 3090 (used)Single GPU | Fits | ~$2,944 + unpriced parts | 162 tok/smedian of 3 runs |
| RTX 3090 (used) + 128 GB DDR5 on AM5 (MoE offload)Hybrid GPU + RAM | Fits | ~$4,833 + unpriced parts | 162 tok/smedian of 3 runs |
| Apple Mac Studio (M5 Max, 40-core GPU)Apple unified memory | Fitsassumed limitfrom 48 GB | ~$3,099–$5,099 | 132 tok/smedian of 4 runs |
| Apple MacBook Pro 14-inch (M5 Max, 40-core GPU)Apple unified memory | Fitsassumed limitfrom 48 GB | ~$4,699–$6,699 | 132 tok/smedian of 4 runs |
| Apple MacBook Pro 16-inch (M5 Max, 40-core GPU)Apple unified memory | Fitsassumed limitfrom 48 GB | ~$4,999–$6,999 | 132 tok/smedian of 4 runs |
| Apple Mac Studio (M5 Max, 32-core GPU)Apple unified memory | Fitsassumed limit | ~$2,499 | 105 tok/smedian of 3 runs |
| Apple MacBook Pro 14-inch (M5 Max, 32-core GPU)Apple unified memory | Fitsassumed limit | ~$4,099 | 105 tok/smedian of 3 runs |
| Apple MacBook Pro 16-inch (M5 Max, 32-core GPU)Apple unified memory | Fitsassumed limit | ~$4,399 | 105 tok/smedian of 3 runs |
| NVIDIA DGX SparkNVIDIA GB10 | Fitsassumed limit | ~$4,699 | 83 tok/smedian of 13 runs |
| Dell Pro Precision with GB10NVIDIA GB10 | Fitsassumed limit | ~$5,523 | 83 tok/smedian of 13 runs |
| Apple Mac mini (M5 Pro, 16-core GPU)Apple unified memory | Fitsassumed limitfrom 24 GB | ~$1,699–$2,699 | 80 tok/smedian of 3 runs |
| Apple Mac mini (M5 Pro, 20-core GPU)Apple unified memory | Fitsassumed limitfrom 24 GB | ~$1,899–$2,899 | 80 tok/smedian of 3 runs |
| Framework Desktop (Ryzen AI Max+ 395)AMD unified memory | Fitsassumed limitfrom 64 GB | ~$1,959–$3,449 | 73 tok/smedian of 3 runs |
| GMKtec EVO-X2 (Ryzen AI Max+ 395)AMD unified memory | Fitsassumed limitfrom 64 GB | ~$2,600–$4,000 | 73 tok/smedian of 3 runs |
| Beelink GTR9 Pro (Ryzen AI Max+ 395)AMD unified memory | Fits | ~$4,699 | 73 tok/smedian of 3 runs |
| Minisforum MS-S1 MAX (Ryzen AI Max+ 395)AMD unified memory | Fits | ~$4,899 | 73 tok/smedian of 3 runs |
| Corsair AI Workstation 300 (Ryzen AI Max+ 395)AMD unified memory | Fits | ~$4,800 | 73 tok/smedian of 3 runs |
| HP Z2 Mini G1a (Ryzen AI Max+ PRO 395)AMD unified memory | Fits | ~$4,999 | 73 tok/smedian of 3 runs |
Estimated, not measured
These fit but nobody has published a run we could match. The figure is memory bandwidth divided by the bytes read per token, scaled by the share of that ceiling measured runs reach. How estimates work.
| Machine | Fit | Price | Generation |
|---|---|---|---|
| Dell Pro Precision with GB300NVIDIA workstation | Fits | ~$159,283 | ~1,678–2,143 tok/sestimate |
| 2x RTX 5090Multi-GPU | Fits | ~$4,109 + unpriced parts | ~224–327 tok/sestimate |
| Apple Mac Studio (M5 Ultra, 80-core GPU)Apple unified memory | Fitsassumed limitfrom 96 GB | ~$6,799–$10,799 | ~200–357 tok/sestimate |
| 4x Instinct MI50 32GB (used) on EPYC 7003Multi-GPU | Fits | ~$1,998 + unpriced parts | ~128–187 tok/sestimate |
| 2x Radeon RX 7900 XTXMulti-GPU | Fits | ~$5,637 | ~120–175 tok/sestimate |
| 2x RTX 3090 (used)Multi-GPU | Fits | ~$3,639 + unpriced parts | ~117–171 tok/sestimate |
| 4x RTX 3090 (used) on EPYC 7003, open frameMulti-GPU | Fits | ~$2,178 + unpriced parts | ~117–171 tok/sestimate |
| tiny corp tinybox red flashGPU desktop (prebuilt) | Fits | ~$22,000 | ~80–117 tok/sestimate |
| Framework Desktop (Ryzen AI Max+ PRO 495)AMD unified memory | Fitsassumed limit | ~$6,799 | ~65–82 tok/sestimate |
| 2x Intel Arc Pro B60Multi-GPU | Fits | ~$4,367 | ~57–83 tok/sestimate |
| Framework Desktop (Ryzen AI Max 385)AMD unified memory | Fitsassumed limit | ~$1,269 | ~61–77 tok/sestimate |
Fits, no speed figure
Files
Released with its experts in MXFP4. GGUF files labelled Q2_K to Q8_0 differ only in the small non-expert layers and are all about this size, so only the native weights are listed.
| Quant | Format | Size | Repo |
|---|---|---|---|
| MXFP4 (native) | GGUF | 10.8 GB | unsloth/gpt-oss-20b-GGUF |
| MXFP4-Q8 | MLX | 11.2 GB | mlx-community/gpt-oss-20b-MXFP4-Q8 |
| 8bit | MLX | 20.7 GB | lmstudio-community/gpt-oss-20b-MLX-8bit |
Sources
- huggingface.co: model card: parameters, context length, license and attention layout; release month
- huggingface.co: downloads (last 30 days), license tag, safetensors param count
- huggingface.co: layer count, KV heads and head size
- huggingface.co: GGUF file sizes (shards summed)
- huggingface.co: MLX 8bit safetensors sizes
- huggingface.co: MLX MXFP4-Q8 safetensors sizes