How we got this number
Each generated token reads the whole model from memory once: 40 GB. The NVIDIA DGX Spark (128GB) moves 273 GB/s, so the ceiling is 6.8 tokens/sec. Real machines reach a fraction of that ceiling; we use 70%, an assumption until we get measured runs for this kind of machine. Send us your real numbers and we'll replace the estimate.
Faster machines for Llama 3 70B
| Mac Studio M5 Ultra (96GB) | 25 tok/s | Check price → |
| Mac Studio M5 Ultra (256GB) | 25 tok/s | Check price → |
| Mac Studio M5 Max (40-core GPU, 64GB) | 12 tok/s | Check price → |
| Mac mini M5 Pro (64GB) | 6.1 tok/s | Check price → |
Other models on the NVIDIA DGX Spark (128GB)
| Qwen 3 30B (A3B) | Fast | 56 tok/s |
| Qwen3 Coder 30B (A3B) | Fast | 56 tok/s |
| Qwen 3-VL 30B (A3B) | Fast | 53 tok/s |
| Qwen3.5 35B (A3B) | Fast | 50 tok/s |
| Qwen3.6 35B (A3B) | Fast | 50 tok/s |
| gpt-oss 20B | Fast | 46 tok/s |