Mac Studio M3 Ultra 96GB vs Mac mini M4 Pro 64GB
For running models locally, priced in Canada. At 8k context the first runs 38 models fully in memory, the second 36.
Specifications
| Spec | Mac Studio M3 Ultra 96GB | Mac mini M4 Pro 64GB |
|---|---|---|
| Chip | Apple M3 Ultra (80-core GPU) | Apple M4 Pro (20-core GPU) |
| Memory bus | 1024-bit | 256-bit |
| Memory bandwidth | 819 GB/s | 273 GB/s |
| FP16 compute | 28 TFLOPS | 8.08 TFLOPS |
| Software | Metal | Metal |
| Unified memory | 96 GB | 64 GB |
| GPU can use by default | 72 GB | 48 GB |
| GPU can use, tuned | 88 GB | 56 GB |
| Operating system | macos | macos |
| Processor | M3 Ultra 32-core | M4 Pro 14-core |
| Storage | 1000 GB | 1000 GB |
What each runs
Estimated decode speed with the most precise quant that fits each one est.
| Model | Mac Studio M3 Ultra 96GB | Mac mini M4 Pro 64GB |
|---|---|---|
| Qwen 3.8 27B | 18tok/sQ8_0 | 6tok/sQ8_0 |
| Qwen 3.5 122B-A10B | 66tok/sQ3_K_M | 29tok/sUD-Q2_K_XL |
| Qwen 3.6 27B | 18tok/sQ8_0 | 6tok/sQ8_0 |
| gpt-oss 120B | 93tok/sMXFP4 | won't run |
| Qwen3 Coder Next | 92tok/sQ6_K | 40tok/sQ4_K_M |
| Nemotron 3 Super 120B-A12B | 51tok/sUD-Q3_K_M | won't run |
| Gemma 4 31B | 16tok/sQ8_0 | 5tok/sQ8_0 |
| Mistral Small 4 119B | 70tok/sUD-Q4_K_M | 42tok/sUD-Q2_K_XL |
| Qwen 3.6 35B-A3B | 97tok/sQ8_0 | 32tok/sQ8_0 |
| Qwen 3.5 27B | 18tok/sQ8_0 | 6tok/sQ8_0 |
| Gemma 4 26B-A4B | 70tok/sQ8_0 | 23tok/sQ8_0 |
| GLM 4.5 Air | 25tok/sQ4_K_M | 13tok/sQ2_K |
| Qwen3 Next 80B-A3B | 119tok/sQ6_K | 52tok/sQ4_K_M |
| Qwen 3.5 35B-A3B | 97tok/sQ8_0 | 32tok/sQ8_0 |
| GLM 4.7 Flash | 56tok/sQ8_0 | 19tok/sQ8_0 |
| Devstral Small 2 24B | 20tok/sQ8_0 | 7tok/sQ8_0 |
| Nemotron 3.5 Lightning 30B-A3B | 93tok/sUD-Q6_K_XL | 31tok/sUD-Q6_K_XL |
| gpt-oss 20B | 121tok/sMXFP4 | 40tok/sMXFP4 |
| Kimi Linear 48B-A3B | 97tok/sQ8_0 | 41tok/sQ6_K |
| Qwen3 Coder 30B-A3B | 80tok/sQ8_0 | 27tok/sQ8_0 |
| Qwen3 32B | 14tok/sQ8_0 | 5tok/sQ8_0 |
| Gemma 4 12B | 41tok/sQ8_0 | 14tok/sQ8_0 |
| Qwen 3.5 9B | 54tok/sQ8_0 | 18tok/sQ8_0 |
| Mistral Small 3.2 24B | 20tok/sQ8_0 | 7tok/sQ8_0 |
| Gemma 3 27B | 18tok/sQ8_0 | 6tok/sQ8_0 |
| DeepSeek R1 Distill Qwen 32B | 14tok/sQ8_0 | 5tok/sQ8_0 |
| Llama 3.3 70B | 10tok/sQ5_K_M | 4tok/sQ4_K_M |
| Ministral 3 14B | 34tok/sQ8_0 | 11tok/sQ8_0 |
| Qwen3 14B | 31tok/sQ8_0 | 10tok/sQ8_0 |
| DeepSeek R1 0528 Qwen3 8B | 54tok/sQ8_0 | 18tok/sQ8_0 |
| Qwen3 8B | 54tok/sQ8_0 | 18tok/sQ8_0 |
| Phi-4 14B | 31tok/sQ8_0 | 10tok/sQ8_0 |
| Qwen 3.5 4B | 111tok/sQ8_0 | 37tok/sQ8_0 |
| Ministral 3 8B | 52tok/sQ8_0 | 17tok/sQ8_0 |
| Granite 4.2 8B | 50tok/sQ8_0 | 17tok/sQ8_0 |
| Gemma 4 E4B | 63tok/sQ8_0 | 21tok/sQ8_0 |
| LFM2.5 8B-A1B | 218tok/sQ8_0 | 73tok/sQ8_0 |
| Llama 3.1 8B | 69tok/sQ6_K | 23tok/sQ6_K |
| DeepSeek V4 Flash | won't run | won't run |
| GLM 5.3 Flash | won't run | won't run |
| Qwen 3.8 Flash Next | won't run | won't run |
| MiniMax M2.7 | won't run | won't run |