Which machines run gpt-oss-120b?
openai/gpt-oss-120b · 116.8B parameters, 5.1B active (MoE) · license see model card · on Hugging Face since 2025-08-04
Memory for weights (estimate)
MXFP4 (native): ≈63.4 GB
Q8_0: ≈124.1 GB
Computed from parameter count × bits per weight. Plus context.
Fits at 4-bit
16 of 24 tracked devices
Measured results
15 public tok/s results
Speeds come from different engines and quants.
| Device | Memory | MXFP4 (native) | Q8_0 | tok/s (measured) | Quant, engine | $ per tok/s | € per tok/s |
|---|---|---|---|---|---|---|---|
| Mac Studio M5 Ultra 80-core GPU, 256GB / 1TB | 256GB | fits | fits | 130 | MXFP4, LM Studio, stated in video | $83 | €96 |
| Mac Studio M5 Max 40-core GPU, 128GB / 512GB | 128GB | fits | doesn't fit | 79.1 | MXFP4, Ollama 0.34.1 | $64 | €75 |
| MacBook Pro 16" M5 Max 40-core GPU, 128GB / 2TB | 128GB | fits | doesn't fit | 79.1 | MXFP4, Ollama 0.34.1 | $88 | €101 |
| NVIDIA DGX Spark (Founders Edition) | 128GB | fits | doesn't fit | 58.72 | MXFP4, llama.cpp CUDA llama-bench tg32 | $80 | €71 |
| ASUS Ascent GX10 (GB10) | 128GB | fits | doesn't fit | 58.72 | MXFP4, llama.cpp CUDA llama-bench tg32 | $102 | €91 |
| Dell Pro Max / Pro Precision with GB10, 128GB / 2TB | 128GB | fits | doesn't fit | 58.72 | MXFP4, llama.cpp CUDA llama-bench tg32 | $101 | €89 |
| MSI EdgeXpert MS-C931 (GB10) | 128GB | fits | doesn't fit | 58.72 | MXFP4, llama.cpp CUDA llama-bench tg32 | – | – |
| Lenovo ThinkStation PGX (GB10) | 128GB | fits | doesn't fit | 58.72 | MXFP4, llama.cpp CUDA llama-bench tg32 | – | – |
| GMKtec EVO-X2 128GB / 1TB | 128GB | fits | doesn't fit | 52.74 | MXFP4, llama.cpp Vulkan llama-bench tg32 | $66 | €63 |
| GMKtec EVO-X2 128GB / 2TB | 128GB | fits | doesn't fit | 52.74 | MXFP4, llama.cpp Vulkan llama-bench tg32 | $69 | €65 |
| GMKtec EVO-X3 128GB / 2TB | 128GB | fits | doesn't fit | 52.74 | MXFP4, llama.cpp Vulkan llama-bench tg32 | $72 | €67 |
| Minisforum MS-S1 MAX 128GB / 2TB | 128GB | fits | doesn't fit | 52.74 | MXFP4, llama.cpp Vulkan llama-bench tg32 | $74 | €65 |
| Beelink GTR9 Pro 128GB / 2TB | 128GB | fits | doesn't fit | 52.74 | MXFP4, llama.cpp Vulkan llama-bench tg32 | $82 | €73 |
| Framework Desktop 128GB (DIY) | 128GB | fits | doesn't fit | 52.74 | MXFP4, llama.cpp Vulkan llama-bench tg32 | $65 | €74 |
| HP Z2 Mini G1a (Ryzen AI Max+ PRO 395) | 128GB | fits | doesn't fit | 52.74 | MXFP4, llama.cpp Vulkan llama-bench tg32 | – | – |
| Mac Studio M5 Ultra 64-core GPU, 96GB / 1TB | 96GB | fits | doesn't fit | no public result | – | – | |
| GMKtec EVO-X2 64GB / 1TB | 64GB | doesn't fit | doesn't fit | – | – | – | |
| Minisforum MS-S1 MAX 64GB / 2TB | 64GB | doesn't fit | doesn't fit | – | – | – | |
| Framework Desktop 64GB (DIY) | 64GB | doesn't fit | doesn't fit | – | – | – | |
| Mac mini M5 Pro 16-core GPU, 64GB / 512GB | 64GB | doesn't fit | doesn't fit | – | – | – | |
| Mac mini M5 Pro 20-core GPU, 64GB / 512GB | 64GB | doesn't fit | doesn't fit | – | – | – | |
| NVIDIA RTX 5090 32GB (new) card, needs host PC | 32GB | doesn't fit | doesn't fit | – | – | – | |
| NVIDIA RTX 4090 24GB (used) card, needs host PC | 24GB | doesn't fit | doesn't fit | – | – | – | |
| NVIDIA RTX 3090 24GB (used) card, needs host PC | 24GB | doesn't fit | doesn't fit | – | – | – |
Fit is an estimate: weights × 1.15 + 2 GB must fit in usable memory (VRAM − 1 GB for cards, 85% of unified memory for machines). tok/s is text generation from the linked public source for that chip; "no public result" means we found none, never a guess. Why this model is listed: Top-5 text-generation by likes. Native MXFP4.