Hardware for running open-weight models
Each page grades every open-weight model on one device. Its memory, bandwidth and compute are its maker’s own figures, cited.
GeForce RTX 50
GeForce RTX 40
GeForce RTX 30
Workstation
Data center
Radeon
Instinct
Arc
Apple silicon
Apple M4 16 GBApple M4 24 GBApple M4 32 GBApple M5 16 GBApple M5 24 GBApple M5 32 GBApple M6 (16GB) 16 GBApple M6 (24GB or 32GB) 24 GBApple M6 (24GB or 32GB) 32 GBApple M4 Pro 24 GBApple M4 Pro 48 GBApple M4 Pro 64 GBApple M5 Pro (16-core GPU) 24 GBApple M5 Pro (16-core GPU) 48 GBApple M5 Pro (20-core GPU) 24 GBApple M5 Pro (20-core GPU) 48 GBApple M5 Pro (20-core GPU) 64 GBApple M4 Max (32-core GPU) 36 GBApple M5 Max (32-core GPU) 36 GBApple M4 Max (40-core GPU) 48 GBApple M4 Max (40-core GPU) 64 GBApple M4 Max (40-core GPU) 128 GBApple M5 Max (40-core GPU) 48 GBApple M5 Max (40-core GPU) 64 GBApple M5 Max (40-core GPU) 128 GBApple M3 Ultra 96 GBApple M3 Ultra 256 GBApple M5 Ultra (64-core GPU) 96 GBApple M5 Ultra (64-core GPU) 256 GBApple M5 Ultra (80-core GPU) 96 GBApple M5 Ultra (80-core GPU) 256 GBApple M5 Ultra (80-core GPU) 512 GB