How we got this number
Each generated token reads the whole model from memory once: 43 GB. The Mac Studio M5 Max (40-core GPU, 64GB) moves 614 GB/s, so the ceiling is 14 tokens/sec. Real machines reach a fraction of that ceiling; we use 80%, calibrated against a measured Mac run. Send us your real numbers and we'll replace the estimate.
Faster machines for Llama 3.3 70B
| Mac Studio M5 Ultra (96GB) | 23 tok/s | Check price → |
| Mac Studio M5 Ultra (256GB) | 23 tok/s | Check price → |
| Mac mini M5 Pro (64GB) | 5.7 tok/s | Check price → |
| NVIDIA DGX Spark (128GB) | 4.4 tok/s | Check price → |
Other models on the Mac Studio M5 Max (40-core GPU, 64GB)
| Qwen 3 30B (A3B) | Fast | 120+ tok/s |
| Qwen3.5 35B (A3B) | Fast | 120+ tok/s |
| Qwen3 Coder 30B (A3B) | Fast | 120+ tok/s |
| Qwen3.6 35B (A3B) | Fast | 120+ tok/s |
| Qwen 3-VL 30B (A3B) | Fast | 120+ tok/s |
| gpt-oss 20B | Fast | 119 tok/s |