Mac Studio M3 Ultra 256GB vs Mac Studio M4 Max 128GB
For running models locally, priced in Germany. At 128k context the first runs 42 models fully in memory, the second 39.
Specifications
| Spec | Mac Studio M3 Ultra 256GB | Mac Studio M4 Max 128GB |
|---|---|---|
| Chip | Apple M3 Ultra (80-core GPU) | Apple M4 Max (40-core GPU) |
| Memory bus | 1024-bit | 512-bit |
| Memory bandwidth | 819 GB/s | 546 GB/s |
| FP16 compute | 28 TFLOPS | 16.16 TFLOPS |
| Software | Metal | Metal |
| Unified memory | 256 GB | 128 GB |
| GPU can use by default | 192 GB | 96 GB |
| GPU can use, tuned | 248 GB | 120 GB |
| Operating system | macos | macos |
| Processor | M3 Ultra 32-core | M4 Max 16-core |
| Storage | 2000 GB | 1000 GB |
What each runs
Estimated decode speed with the most precise quant that fits each one est.
| Model | Mac Studio M3 Ultra 256GB | Mac Studio M4 Max 128GB |
|---|---|---|
| DeepSeek V4 Flash | 14tok/sUD-Q4_K_XL | won't run |
| GLM 5.3 Flash | 18tok/sUD-Q4_K_XL | won't run |
| Qwen 3.8 Flash Next | 5tok/sQ8_0 | 6tok/sUD-Q3_K_XL |
| Qwen 3.8 27B | 14tok/sQ8_0 | 9tok/sQ8_0 |
| MiniMax M2.7 | 8tok/sUD-Q5_K_M | won't run |
| Qwen 3.5 122B-A10B | 23tok/sQ8_0 | 20tok/sQ5_K_M |
| Qwen 3.6 27B | 14tok/sQ8_0 | 10tok/sQ8_0 |
| Qwen3 Coder Next | 43tok/sQ8_0 | 29tok/sQ8_0 |
| gpt-oss 120B | 40tok/sMXFP4 | 27tok/sMXFP4 |
| Nemotron 3 Super 120B-A12B | 23tok/sQ8_0 | 23tok/sUD-Q4_K_M |
| Gemma 4 31B | 12tok/sQ8_0 | 8tok/sQ8_0 |
| Mistral Small 4 119B | 30tok/sQ8_0 | 26tok/sUD-Q5_K_M |
| Qwen 3.6 35B-A3B | 55tok/sQ8_0 | 37tok/sQ8_0 |
| Qwen 3.5 27B | 14tok/sQ8_0 | 10tok/sQ8_0 |
| Gemma 4 26B-A4B | 45tok/sQ8_0 | 30tok/sQ8_0 |
| GLM 4.5 Air | 8tok/sQ8_0 | 6tok/sQ4_K_M |
| Qwen 3.5 35B-A3B | 55tok/sQ8_0 | 37tok/sQ8_0 |
| Qwen3 Next 80B-A3B | 50tok/sQ8_0 | 33tok/sQ8_0 |
| GLM 4.7 Flash | 26tok/sQ8_0 | 17tok/sQ8_0 |
| Devstral Small 2 24B | 11tok/sQ8_0 | 8tok/sQ8_0 |
| Nemotron 3.5 Lightning 30B-A3B | 76tok/sUD-Q6_K_XL | 51tok/sUD-Q6_K_XL |
| Kimi Linear 48B-A3B | 75tok/sQ8_0 | 50tok/sQ8_0 |
| gpt-oss 20B | 57tok/sMXFP4 | 38tok/sMXFP4 |
| Qwen3 Coder 30B-A3B | 20tok/sQ8_0 | 13tok/sQ8_0 |
| Qwen3 32B | 8tok/sQ8_0 | 5tok/sQ8_0 |
| Gemma 4 12B | 35tok/sQ8_0 | 23tok/sQ8_0 |
| Qwen 3.5 9B | 38tok/sQ8_0 | 26tok/sQ8_0 |
| Gemma 3 27B | 13tok/sQ8_0 | 9tok/sQ8_0 |
| Mistral Small 3.2 24B | 11tok/sQ8_0 | 8tok/sQ8_0 |
| DeepSeek R1 Distill Qwen 32B | 8tok/sQ8_0 | 5tok/sQ8_0 |
| Llama 3.3 70B | 5tok/sQ8_0 | 4tok/sQ5_K_M |
| Ministral 3 14B | 15tok/sQ8_0 | 10tok/sQ8_0 |
| Qwen3 14B | 14tok/sQ8_0 | 10tok/sQ8_0 |
| DeepSeek R1 0528 Qwen3 8B | 19tok/sQ8_0 | 13tok/sQ8_0 |
| Qwen3 8B | 19tok/sQ8_0 | 13tok/sQ8_0 |
| Phi-4 14B | 13tok/sQ8_0 | 8tok/sQ8_0 |
| Qwen 3.5 4B | 60tok/sQ8_0 | 40tok/sQ8_0 |
| Ministral 3 8B | 20tok/sQ8_0 | 13tok/sQ8_0 |
| Granite 4.2 8B | 17tok/sQ8_0 | 12tok/sQ8_0 |
| Gemma 4 E4B | 44tok/sQ8_0 | 30tok/sQ8_0 |
| LFM2.5 8B-A1B | 43tok/sQ8_0 | 28tok/sQ8_0 |
| Llama 3.1 8B | 22tok/sQ6_K | 15tok/sQ6_K |