Best Local AI Models by Device

Pick a machine to see which local models it runs, how much memory each one needs, and the engine's speed estimate: the 5 Mac lines Apple sells today, 11 iPhones and 30 graphics cards and AI PCs.

Compare two machines
Mac
5
iPhone
11
NVIDIA GPU
25
AMD
5
Apple Silicon

Mac

Apple list prices checked 2026-11-16 · US store
On-device
VRAM-bound

Speeds are engine estimates for Qwen3 8B at Q4, not measurements. Table 1 and Table 2 give each card's typical error.

Table 1Specs

NVIDIA GPU

25 cards
GPUTierArchMemoryBandwidthSpeed
RTX 4060BudgetAda Lovelace8 GB GDDR6272 GB/sest. 40 tok/stypical error 5%
RTX 3060BudgetAmpere12 GB GDDR6360 GB/sest. 50 tok/stypical error 5%
RTX 3070BudgetAmpere8 GB GDDR6448 GB/sest. 61 tok/stypical error 5%
RTX 4060 TiBudgetAda Lovelace16 GB GDDR6288 GB/sest. 43 tok/stypical error 5%
RTX 5060 TiBudgetBlackwell16 GB GDDR7448 GB/sest. 68 tok/stypical error 5%
RTX 3060 TiBudgetAmpere8 GB GDDR6448 GB/sest. 61 tok/stypical error 5%
RTX 2070BudgetTuring8 GB GDDR6448 GB/sest. 54 tok/stypical error 5%
RTX 5060BudgetBlackwell8 GB GDDR7448 GB/sest. 68 tok/stypical error 5%
RTX 4070Mid-RangeAda Lovelace12 GB GDDR6X504 GB/sest. 73 tok/stypical error 5%
RTX 4070 SUPERMid-RangeAda Lovelace12 GB GDDR6X504 GB/sest. 73 tok/stypical error 5%
RTX 5070Mid-RangeBlackwell12 GB GDDR7672 GB/sest. 98 tok/stypical error 5%
RTX 5070 TiMid-RangeBlackwell16 GB GDDR7896 GB/sest. 125 tok/stypical error 5%
RTX 4070 Ti SUPERMid-RangeAda Lovelace16 GB GDDR6X672 GB/sest. 94 tok/stypical error 5%
RTX 4070 TiMid-RangeAda Lovelace12 GB GDDR6X504 GB/sest. 73 tok/stypical error 5%
RTX 3080 10GBMid-RangeAmpere10 GB GDDR6X760 GB/sest. 99 tok/stypical error 5%
Tesla P100Mid-RangePascal16 GB HBM2732 GB/sest. 42 tok/stypical error 5%
Tesla T4Mid-RangeTuring16 GB GDDR6300 GB/sest. 37 tok/stypical error 5%
RTX 4080 SUPERHigh-EndAda Lovelace16 GB GDDR6X736 GB/sest. 102 tok/stypical error 5%
RTX 5080High-EndBlackwell16 GB GDDR7960 GB/sest. 132 tok/stypical error 5%
RTX 3090High-EndAmpere24 GB GDDR6X936 GB/sest. 117 tok/stypical error 5%
RTX 4090High-EndAda Lovelace24 GB GDDR6X1008 GB/sest. 133 tok/stypical error 5%
RTX 4080High-EndAda Lovelace16 GB GDDR6X716.8 GB/sest. 100 tok/stypical error 5%
RTX 3090 TiHigh-EndAmpere24 GB GDDR6X1008 GB/sest. 124 tok/stypical error 5%
RTX 5090UltraBlackwell32 GB GDDR71792 GB/sest. 213 tok/stypical error 5%
RTX PRO 6000UltraBlackwell96 GB GDDR71792 GB/sest. 213 tok/stypical error 5%
Table 2Radeon GPUs + Ryzen AI APUs

AMD

5 devices
DeviceTypeMemoryMax modelBandwidthSpeed
RX 7900 XTXRDNA 324 GB GDDR6~33B dense960 GB/sest. 104 tok/stypical error 10%
RX 7900 XTRDNA 320 GB GDDR6~30B dense800 GB/sest. 91 tok/stypical error 10%
RX 9070 XTRDNA 416 GB GDDR6~14B dense640 GB/sest. 79 tok/stypical error 10%
RX 9060 XTRDNA 416 GB GDDR6~14B dense320 GB/sest. 45 tok/stypical error 10%
Ryzen AI Max+ 395Strix Halo110 GB unified~33B dense, ~125B MoE256 GB/sest. 32 tok/stypical error 10%

Getting started guides

Not sure which one to buy?

Put any two machines side by side, a Mac against a Mac or a Mac against a graphics card: usable memory, the models each one runs, and the estimated speed of each.