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 4 Mac lines Apple sells today, 11 iPhones and 19 graphics cards and AI PCs.
Apple Silicon
Mac
Apple list prices checked 2026-10-25 · US storeMacBook Air
M5 · 16 to 32 GBFrom $1,299Apple list price · US · source- Top pick at 16 GB
- Qwen3.5 9B Instruct est. 26 tok/s
- Largest that runs
- Qwen3.6 35B-A3B 35B on 32 GB
MacBook Pro
M5 · M5 Pro · M5 Max · 16 to 128 GBFrom $1,999Apple list price · US · source- Top pick at 16 GB
- Qwen3.5 9B Instruct est. 26 tok/s
- Largest that runs
- Qwen3.8-Flash-Next (IQ4_XS) 125B on 128 GB
Mac Mini
M6 · M5 Pro · 16 to 64 GBFrom $899Apple list price · US · source- Top pick at 16 GB
- Qwen3.5 9B Instruct est. 28 tok/s
- Largest that runs
- Kolibri 1 (IQ4_XS) 78.1B on 64 GB
Mac Studio
M5 Max · M5 Ultra · 36 to 256 GBFrom $2,499Apple list price · US · source- Top pick at 36 GB
- Qwen3.6 35B-A3B est. 128 tok/s
- Largest that runs
- Qwen3 235B A22B 235B on 256 GB
On-device
iPhone
iPhone guide →iPhone 17 family
iPhone 16 family
iPhone 15 family
VRAM-bound
Graphics cards
Compare a card with a Mac →8 GB
Largest model that runs: 9B12 GB
Largest model that runs: 14B16 GB
Largest model that runs: 24B20 to 32 GB
Largest model that runs: 30B to 40BOver 32 GB
Largest model that runs: 124B to 125B
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
16 cards| GPU | Tier | Arch | Memory | Bandwidth | Speed |
|---|---|---|---|---|---|
| RTX 4060 | Budget | Ada Lovelace | 8 GB GDDR6 | 272 GB/s | est. 40 tok/stypical error 5% |
| RTX 3060 | Budget | Ampere | 12 GB GDDR6 | 360 GB/s | est. 50 tok/stypical error 5% |
| RTX 3070 | Budget | Ampere | 8 GB GDDR6 | 448 GB/s | est. 61 tok/stypical error 5% |
| RTX 4060 Ti | Budget | Ada Lovelace | 16 GB GDDR6 | 288 GB/s | est. 43 tok/stypical error 5% |
| RTX 5060 Ti | Budget | Blackwell | 16 GB GDDR7 | 448 GB/s | est. 68 tok/stypical error 5% |
| RTX 4070 | Mid-Range | Ada Lovelace | 12 GB GDDR6X | 504 GB/s | est. 73 tok/stypical error 5% |
| RTX 4070 SUPER | Mid-Range | Ada Lovelace | 12 GB GDDR6X | 504 GB/s | est. 73 tok/stypical error 5% |
| RTX 5070 | Mid-Range | Blackwell | 12 GB GDDR7 | 672 GB/s | est. 98 tok/stypical error 5% |
| RTX 5070 Ti | Mid-Range | Blackwell | 16 GB GDDR7 | 896 GB/s | est. 125 tok/stypical error 5% |
| RTX 4070 Ti SUPER | Mid-Range | Ada Lovelace | 16 GB GDDR6X | 672 GB/s | est. 94 tok/stypical error 5% |
| RTX 4080 SUPER | High-End | Ada Lovelace | 16 GB GDDR6X | 736 GB/s | est. 102 tok/stypical error 5% |
| RTX 5080 | High-End | Blackwell | 16 GB GDDR7 | 960 GB/s | est. 132 tok/stypical error 5% |
| RTX 3090 | High-End | Ampere | 24 GB GDDR6X | 936 GB/s | est. 117 tok/stypical error 5% |
| RTX 4090 | High-End | Ada Lovelace | 24 GB GDDR6X | 1008 GB/s | est. 133 tok/stypical error 5% |
| RTX 5090 | Ultra | Blackwell | 32 GB GDDR7 | 1792 GB/s | est. 213 tok/stypical error 5% |
| RTX PRO 6000 | Ultra | Blackwell | 96 GB GDDR7 | 1792 GB/s | est. 213 tok/stypical error 5% |
Table 2Radeon GPUs + Ryzen AI APUs
AMD
3 devices| Device | Type | Memory | Max model | Bandwidth | Speed |
|---|---|---|---|---|---|
| RX 7900 XTX | RDNA 3 | 24 GB GDDR6 | 32B parameter models | 960 GB/s | est. 104 tok/stypical error 10% |
| RX 7900 XT | RDNA 3 | 20 GB GDDR6 | 27B parameter models | 800 GB/s | est. 91 tok/stypical error 10% |
| Ryzen AI Max+ 395 | Strix Halo | 110 GB unified | 120B parameter models | 256 GB/s | est. 32 tok/stypical error 10% |
Fig. 1Max model size by memory
Largest dense Q4 model that fits each memory tier
8 GB9B
12 GB12B
16 GB14B
24 GB29.3B
32 GB35B
36 GB35B
48 GB35B
64 GB70B
72 GB70B
96 GB123B
128 GB123B
192 GB123B
256 GB123B
512 GB405B
billion parametersGetting 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.