GPU directory
Which GPUs run local AI well?
VRAM capacity is usually the first constraint in local AI, and memory bandwidth decides how fast tokens are produced. Browse cards by vendor to see what model classes each one can realistically host.
NVIDIA
NVIDIA GeForce RTX 4050 Laptop GPU
6 GB3B (Q8/FP16) · 7B (Q4/Q5) · 8B (Q4)
Ada LovelaceLaptop
View GPU page NVIDIA GeForce RTX 4060
8 GB7B (Q5/Q6) · 8B (Q4/Q5) · 12B (Q3/Q4) · 14B (Q3)
Ada Lovelace
View GPU page NVIDIA GeForce RTX 4060 Ti
8 GB7B (Q6) · 8B (Q5) · 14B (Q4, 16 GB variant only)
Ada Lovelace
View GPU page NVIDIA GeForce RTX 4070
12 GB7B (Q8) · 8B (Q6/Q8) · 12B (Q5) · 14B (Q4/Q5)
Ada Lovelace
View GPU page NVIDIA GeForce RTX 4080
16 GB7B (FP16) · 14B (Q6/Q8) · 32B (Q3/Q4)
Ada Lovelace
View GPU page NVIDIA GeForce RTX 4090
24 GB14B (FP16) · 32B (Q4/Q5) · 70B (Q2/Q3 with offload)
Ada Lovelace
View GPU page NVIDIA GeForce RTX 3060 12 GB
12 GB7B (Q6/Q8) · 8B (Q5) · 12B (Q4/Q5) · 13B (Q4)
Ampere
View GPU page AMD
Apple
Apple M4 Pro
27 GB7B (FP16) · 14B (Q8) · 32B (Q4)
Apple SiliconLaptopUnified memory
View GPU page Apple M2 (8 GB unified memory)
8 GB1B (Q4+) · 3B (Q4)
Apple SiliconLaptopUnified memory
View GPU page Apple M3 Pro (18 GB unified memory)
18 GB7B (Q6/Q8) · 14B (Q4/Q5)
Apple SiliconLaptopUnified memory
View GPU page Apple M4 Max (36 GB+ unified memory)
36 GB14B (FP16) · 32B (Q5/Q6) · 70B (Q2/Q3)
Apple SiliconLaptopUnified memory
View GPU page Want a personalized answer?
Enter your GPU, RAM and CPU once — the checker grades an entire model catalogue against your exact profile.