Every M chip for local AI.
Memory decides which models fit; bandwidth decides how fast they write. Here is every Apple Silicon chip on both, worked out from Apple's own figures.
| Chip | Unified memory | Memory bandwidth | Biggest model at 4-bit | 8B model, ceiling | Found in |
|---|---|---|---|---|---|
| M1 | 16 GB | 68 GB/s | about 21B | 15 tokens/s | MacBook Air (2020), 13-inch MacBook Pro (2020), Mac mini (2020), iMac (2021) |
| M1 Pro | 32 GB | 200 GB/s | about 42B | 44 tokens/s | 14- and 16-inch MacBook Pro (2021) |
| M1 Max | 64 GB | 400 GB/s | about 85B | 89 tokens/s | 14- and 16-inch MacBook Pro (2021), Mac Studio (2022) |
| M1 Ultra | 128 GB | 800 GB/s | about 170B | 178 tokens/s | Mac Studio (2022) |
| M2 | 24 GB | 100 GB/s | about 32B | 22 tokens/s | MacBook Air (2022, 2023), 13-inch MacBook Pro (2022), Mac mini (2023) |
| M2 Pro | 32 GB | 200 GB/s | about 42B | 44 tokens/s | 14- and 16-inch MacBook Pro (2023), Mac mini (2023) |
| M2 Max | 96 GB | 400 GB/s | about 128B | 89 tokens/s | 14- and 16-inch MacBook Pro (2023), Mac Studio (2023) |
| M2 Ultra | 192 GB | 800 GB/s | about 256B | 178 tokens/s | Mac Studio (2023), Mac Pro (2023) |
| M3 | 24 GB | 100 GB/s | about 32B | 22 tokens/s | MacBook Air (2024), 14-inch MacBook Pro (2023), iMac (2023) |
| M3 Pro | 36 GB | 150 GB/s | about 48B | 33 tokens/s | 14- and 16-inch MacBook Pro (2023) |
| M3 Max | 128 GB | 400 GB/s | about 170B | 89 tokens/s | 14- and 16-inch MacBook Pro (2023) |
| M3 Ultra | 512 GB | 819 GB/s | about 682B | 182 tokens/s | Mac Studio (2025) |
| M4 | 32 GB | 120 GB/s | about 42B | 27 tokens/s | MacBook Air (2025), 14-inch MacBook Pro (2024), Mac mini (2024), iMac (2024) |
| M4 Pro | 64 GB | 273 GB/s | about 85B | 61 tokens/s | Mac mini (2024), 14- and 16-inch MacBook Pro (2024) |
| M4 Max | 128 GB | 546 GB/s | about 170B | 121 tokens/s | 14- and 16-inch MacBook Pro (2024), Mac Studio (2025) |
| M5 | 32 GB | 153 GB/s | about 42B | 34 tokens/s | 14-inch MacBook Pro (2025), MacBook Air (2026) |
| M5 Pro | 64 GB | 307 GB/s | about 85B | 68 tokens/s | 14- and 16-inch MacBook Pro (2026), Mac mini (2026) |
| M5 Max | 128 GB | 614 GB/s | about 170B | 136 tokens/s | 14- and 16-inch MacBook Pro (2026), Mac Studio (2026) |
| M5 Ultra | 512 GB | 1.2 TB/s | about 682B | 267 tokens/s | Mac Studio (2026) |
| M6 | 32 GB | 170 GB/s | about 42B | 38 tokens/s | Mac mini (2026) |
Ceilings are memory bandwidth divided by the bytes read per token; real runs usually reach about 45% to 80% of them. The biggest model assumes three quarters of the memory is usable. Each chip's page lists its sources.
Sources
- Apple: M1apple.com
- llama.cpp: Apple Silicon benchmarksgithub.com
- Apple: M1 Pro and M1 Maxapple.com
- Apple: M1 Ultraapple.com
- Apple: M2apple.com
- Apple: M2 Pro and M2 Maxapple.com
- Apple: M2 Ultraapple.com
- Apple: M3, M3 Pro and M3 Maxapple.com
- Apple: MacBook Pro (M3 Pro or M3 Max) specssupport.apple.com
- Apple: Mac Studio (2025) specssupport.apple.com
- Apple: Mac mini (2024) specssupport.apple.com
- Apple: M4 Pro and M4 Maxapple.com
- Apple: M5apple.com
- Apple: MacBook Air with M5apple.com
- Apple: M5 Pro and M5 Maxapple.com
- Apple: Mac mini (2026) specssupport.apple.com
- Apple: Mac Studio with M5 Max and M5 Ultraapple.com
- Apple: M6 and M5 Ultraapple.com