Mac Studio M3 Ultra 96GB vs Mac Studio M3 Ultra 512GB
For running models locally, priced in Germany. At 32k context the first runs 38 models fully in memory, the second 42.
Specifications
| Spec | Mac Studio M3 Ultra 96GB | Mac Studio M3 Ultra 512GB |
|---|---|---|
| Chip | Apple M3 Ultra (80-core GPU) | Apple M3 Ultra (80-core GPU) |
| Memory bus | 1024-bit | 1024-bit |
| Memory bandwidth | 819 GB/s | 819 GB/s |
| FP16 compute | 28 TFLOPS | 28 TFLOPS |
| Software | Metal | Metal |
| Unified memory | 96 GB | 512 GB |
| GPU can use by default | 72 GB | 384 GB |
| GPU can use, tuned | 88 GB | 504 GB |
| Operating system | macos | macos |
| Processor | M3 Ultra 32-core | M3 Ultra 32-core |
| Storage | 1000 GB | 2000 GB |
What each runs
Estimated decode speed with the most precise quant that fits each one est.
| Model | Mac Studio M3 Ultra 96GB | Mac Studio M3 Ultra 512GB |
|---|---|---|
| DeepSeek V4 Flash | won't run | 23tok/sUD-Q4_K_XL |
| GLM 5.3 Flash | won't run | 12tok/sQ8_0 |
| Qwen 3.8 Flash Next | won't run | 5tok/sQ8_0 |
| Qwen 3.8 27B | 17tok/sQ8_0 | 17tok/sQ8_0 |
| MiniMax M2.7 | won't run | 16tok/sQ8_0 |
| Qwen 3.5 122B-A10B | 58tok/sQ3_K_M | 28tok/sQ8_0 |
| Qwen 3.6 27B | 17tok/sQ8_0 | 17tok/sQ8_0 |
| Qwen3 Coder Next | 78tok/sQ6_K | 64tok/sQ8_0 |
| gpt-oss 120B | 74tok/sMXFP4 | 74tok/sMXFP4 |
| Nemotron 3 Super 120B-A12B | 50tok/sUD-Q3_K_M | 25tok/sQ8_0 |
| Gemma 4 31B | 15tok/sQ8_0 | 15tok/sQ8_0 |
| Mistral Small 4 119B | 62tok/sUD-Q4_K_M | 39tok/sQ8_0 |
| Qwen 3.6 35B-A3B | 84tok/sQ8_0 | 84tok/sQ8_0 |
| Qwen 3.5 27B | 17tok/sQ8_0 | 17tok/sQ8_0 |
| Gemma 4 26B-A4B | 63tok/sQ8_0 | 63tok/sQ8_0 |
| GLM 4.5 Air | 21tok/sQ3_K_M | 13tok/sQ8_0 |
| Qwen3 Next 80B-A3B | 97tok/sQ6_K | 80tok/sQ8_0 |
| Qwen 3.5 35B-A3B | 84tok/sQ8_0 | 84tok/sQ8_0 |
| GLM 4.7 Flash | 45tok/sQ8_0 | 45tok/sQ8_0 |
| Devstral Small 2 24B | 18tok/sQ8_0 | 18tok/sQ8_0 |
| Nemotron 3.5 Lightning 30B-A3B | 89tok/sUD-Q6_K_XL | 89tok/sUD-Q6_K_XL |
| gpt-oss 20B | 99tok/sMXFP4 | 99tok/sMXFP4 |
| Kimi Linear 48B-A3B | 91tok/sQ8_0 | 91tok/sQ8_0 |
| Qwen3 Coder 30B-A3B | 50tok/sQ8_0 | 50tok/sQ8_0 |
| Qwen3 32B | 12tok/sQ8_0 | 12tok/sQ8_0 |
| Gemma 4 12B | 39tok/sQ8_0 | 39tok/sQ8_0 |
| Qwen 3.5 9B | 50tok/sQ8_0 | 50tok/sQ8_0 |
| Mistral Small 3.2 24B | 18tok/sQ8_0 | 18tok/sQ8_0 |
| Gemma 3 27B | 17tok/sQ8_0 | 17tok/sQ8_0 |
| DeepSeek R1 Distill Qwen 32B | 12tok/sQ8_0 | 12tok/sQ8_0 |
| Llama 3.3 70B | 9tok/sQ5_K_M | 6tok/sQ8_0 |
| Ministral 3 14B | 27tok/sQ8_0 | 27tok/sQ8_0 |
| Qwen3 14B | 25tok/sQ8_0 | 25tok/sQ8_0 |
| DeepSeek R1 0528 Qwen3 8B | 39tok/sQ8_0 | 39tok/sQ8_0 |
| Qwen3 8B | 39tok/sQ8_0 | 39tok/sQ8_0 |
| Phi-4 14B | 24tok/sQ8_0 | 24tok/sQ8_0 |
| Qwen 3.5 4B | 95tok/sQ8_0 | 95tok/sQ8_0 |
| Ministral 3 8B | 39tok/sQ8_0 | 39tok/sQ8_0 |
| Granite 4.2 8B | 36tok/sQ8_0 | 36tok/sQ8_0 |
| Gemma 4 E4B | 58tok/sQ8_0 | 58tok/sQ8_0 |
| LFM2.5 8B-A1B | 119tok/sQ8_0 | 119tok/sQ8_0 |
| Llama 3.1 8B | 49tok/sQ6_K | 49tok/sQ6_K |