Apple M2 Ultra vs Apple M3 Ultra for local AI
Both fit the same models we track, up to Qwen3 235B-A22B. They generate at about the same speed, since their memory bandwidth is close (800 and 819 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 M2 Ultra | Apple M3 Ultra | |
|---|---|---|
| Memory | 64 or 128 or 192 GB unified memory | 96 or 256 or 512 GB unified memory |
| Usable for models | About 144 GB of 192 GB | About 384 GB of 512 GB |
| Bandwidth | 800 GB/s | 819 GB/s |
| Biggest model that fits | Qwen3 235B-A22B | Qwen3 235B-A22B |
| Best everyday pick | Qwen3 235B-A22B, ~21 tok/s | Qwen3 235B-A22B, ~22 tok/s |
Every model, side by side
Estimated generation speed at the largest memory size each was sold in (192 GB and 512 GB).
| Model | Apple M2 Ultra | Apple M3 Ultra |
|---|---|---|
| Qwen3 4B About 4 GB | ~180 tok/s Runs great | ~180 tok/s Runs great |
| Qwen3 8B About 7 GB | ~85 tok/s Runs great | ~87 tok/s Runs great |
| Gemma 3 12B About 10 GB | ~54 tok/s Runs great | ~56 tok/s Runs great |
| Qwen3 14B About 11 GB | ~47 tok/s Runs great | ~48 tok/s Runs great |
| gpt-oss 20B About 16 GB | ~140 tok/s Runs great | ~140 tok/s Runs great |
| Gemma 3 27B About 19 GB | ~26 tok/s Usable | ~26 tok/s Usable |
| Qwen3 30B-A3B About 21 GB | ~140 tok/s Runs great | ~140 tok/s Runs great |
| Qwen3 32B About 22 GB | ~22 tok/s Usable | ~23 tok/s Usable |
| Llama 3.3 70B About 45 GB | ~10 tok/s Usable | ~10 tok/s Usable |
| gpt-oss 120B About 67 GB | ~97 tok/s Runs great | ~99 tok/s Runs great |
| Qwen3 235B-A22B About 144 GB | ~21 tok/s Usable | ~22 tok/s Usable |
| DeepSeek R1 671B About 406 GB | Needs 406 GB | Needs 406 GB |
They're close for local AI, so pick on price. 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.