Apple M3 Max vs Apple M5 Max for local AI
Both fit the same models we track, up to gpt-oss 120B. The Apple M5 Max generates about 15% faster on any model both can run, because it has 460 GB/s of memory bandwidth to the Apple M3 Max's 400 GB/s.
Memory and bandwidth
Memory decides which models fit. Bandwidth decides how fast they generate, because every token reads the model's active weights once.
| Apple M3 Max | Apple M5 Max | |
|---|---|---|
| Memory | 36 or 48 or 64 or 96 or 128 GB unified memory | 36 or 48 or 64 or 128 GB unified memory |
| Usable for models | About 96 GB of 128 GB | About 96 GB of 128 GB |
| Bandwidth | 400 GB/s | 460 GB/s |
| Biggest model that fits | gpt-oss 120B | gpt-oss 120B |
| Best everyday pick | gpt-oss 120B, ~48 tok/s | gpt-oss 120B, ~56 tok/s |
Every model, side by side
Estimated generation speed at the largest memory size each was sold in (128 GB and 128 GB).
| Model | Apple M3 Max | Apple M5 Max |
|---|---|---|
| Qwen3 4B About 4 GB | ~88 tok/s Runs great | ~100 tok/s Runs great |
| Qwen3 8B About 7 GB | ~42 tok/s Runs great | ~49 tok/s Runs great |
| Gemma 3 12B About 10 GB | ~27 tok/s Usable | ~31 tok/s Runs great |
| Qwen3 14B About 11 GB | ~24 tok/s Usable | ~27 tok/s Usable |
| gpt-oss 20B About 16 GB | ~70 tok/s Runs great | ~81 tok/s Runs great |
| Gemma 3 27B About 19 GB | ~13 tok/s Usable | ~15 tok/s Usable |
| Qwen3 30B-A3B About 21 GB | ~70 tok/s Runs great | ~81 tok/s Runs great |
| Qwen3 32B About 22 GB | ~11 tok/s Usable | ~13 tok/s Usable |
| Llama 3.3 70B About 45 GB | ~5 tok/s Slow | ~6 tok/s Slow |
| gpt-oss 120B About 67 GB | ~48 tok/s Runs great | ~56 tok/s Runs great |
| Qwen3 235B-A22B About 144 GB | Needs 144 GB | Needs 144 GB |
| DeepSeek R1 671B About 406 GB | Needs 406 GB | Needs 406 GB |
The Apple M5 Max, for speed. They fit the same models. Price matters as much as either number: check used prices on the builds page, where every parts list has a measured or estimated speed.
Both columns use the same formula: memory bandwidth divided by the weights read per token, calibrated on our own rack, with Ollama's default 4-bit sizes. Real results are usually within about a third. Measured either one? Send us the number.