Mac Studio M3 Ultra 512GB vs NVIDIA RTX PRO 6000 Blackwell
For running models locally, priced in the UK. At 8k context the first runs 93 models fully in memory, the second 75.
In short
The Mac Studio M3 Ultra 512GB has more memory for models: 512 GB against 96 GB. At 8k context the Mac Studio M3 Ultra 512GB runs 93 of the models we track fully in memory and the NVIDIA RTX PRO 6000 Blackwell 75; only the Mac Studio M3 Ultra 512GB holds GLM 5.3 Flash. On the models both run fully in memory, the NVIDIA RTX PRO 6000 Blackwell is about 2.1× as fast as the Mac Studio M3 Ultra 512GB (estimated, median across models). In the UK the Mac Studio M3 Ultra 512GB costs £17,495 at eBay (used) and the NVIDIA RTX PRO 6000 Blackwell £14,499 at eBay (used): the NVIDIA RTX PRO 6000 Blackwell is £2,996 cheaper.
Updated . Speeds are estimates; how they are worked out.
Specifications
| Spec | Mac Studio M3 Ultra 512GB | NVIDIA RTX PRO 6000 Blackwell |
|---|---|---|
| Chip | Apple M3 Ultra (80-core GPU) | – |
| Memory bus | 1024-bit | 512-bit |
| Memory bandwidth | 819 GB/s | 1,792 GB/s |
| FP16 compute | 28 TFLOPS | 504 TFLOPS |
| Software | Metal | CUDA |
| Unified memory | 512 GB | – |
| GPU can use by default | 384 GB | – |
| GPU can use, tuned | 504 GB | – |
| Operating system | macos | – |
| Processor | M3 Ultra 32-core | – |
| Storage | 2000 GB | – |
| GPU | – | RTX PRO 6000 Blackwell |
| Memory | – | 96 GB GDDR7 |
| Board power | – | 600 W |
| PCIe generation | – | Gen 5 |
| PCIe lanes | – | x16 |
What each runs
Estimated decode speed with the most precise quant that fits each one est.
| Model | Mac Studio M3 Ultra 512GB | NVIDIA RTX PRO 6000 Blackwell |
|---|---|---|
| DeepSeek V4 Flash | 27tok/sMXFP4 | 89tok/sUD-Q2_K_XL |
| GLM 5.3 Flash | 12tok/sQ8_0 | offload |
| Qwen 3.8 Flash Next | 5tok/sQ8_0 | 21tok/sUD-Q3_K_XL |
| MiniMax M2.7 | 23tok/sQ8_0 | 117tok/sUD-Q2_K_XL |
| Qwen 3.8 27B | 18tok/sQ8_0 | 37tok/sQ8_0 |
| Qwen 3.5 122B-A10B | 30tok/sQ8_0 | 79tok/sQ6_K |
| Qwen 3.6 27B | 18tok/sQ8_0 | 38tok/sQ8_0 |
| gpt-oss 120B | 93tok/sMXFP4 | 195tok/sMXFP4 |
| Qwen3 Coder Next | 73tok/sQ8_0 | 151tok/sQ8_0 |
| Nemotron 3 Super 120B-A12B | 25tok/sQ8_0 | 69tok/sQ5_K_M |
| Gemma 4 31B | 16tok/sQ8_0 | 33tok/sQ8_0 |
| Mistral Small 4 119B | 42tok/sQ8_0 | 110tok/sUD-Q6_K |
| Qwen 3.6 35B-A3B | 97tok/sQ8_0 | 201tok/sQ8_0 |
| Qwen 3.5 27B | 18tok/sQ8_0 | 38tok/sQ8_0 |
| Gemma 4 26B-A4B | 70tok/sQ8_0 | 147tok/sQ8_0 |
| GLM 4.5 Air | 20tok/sQ8_0 | 48tok/sQ6_K |
| Qwen 3.5 35B-A3B | 97tok/sQ8_0 | 201tok/sQ8_0 |
| Qwen3 Next 80B-A3B | 94tok/sQ8_0 | 196tok/sQ8_0 |
| GLM 4.7 Flash | 56tok/sQ8_0 | 117tok/sQ8_0 |
| Devstral Small 2 24B | 20tok/sQ8_0 | 42tok/sQ8_0 |
| Nemotron 3.5 Lightning 30B-A3B | 93tok/sUD-Q6_K_XL | 195tok/sUD-Q6_K_XL |
| gpt-oss 20B | 121tok/sMXFP4 | 253tok/sMXFP4 |
| Kimi Linear 48B-A3B | 97tok/sQ8_0 | 202tok/sQ8_0 |
| Qwen3 Coder 30B-A3B | 80tok/sQ8_0 | 167tok/sQ8_0 |
| Qwen3 32B | 14tok/sQ8_0 | 30tok/sQ8_0 |
| Gemma 4 12B | 41tok/sQ8_0 | 85tok/sQ8_0 |
| Qwen 3.5 9B | 54tok/sQ8_0 | 113tok/sQ8_0 |
| Mistral Small 3.2 24B | 20tok/sQ8_0 | 42tok/sQ8_0 |
| Gemma 3 27B | 18tok/sQ8_0 | 37tok/sQ8_0 |
| DeepSeek R1 Distill Qwen 32B | 14tok/sQ8_0 | 30tok/sQ8_0 |
| Llama 3.3 70B | 7tok/sQ8_0 | 14tok/sQ8_0 |
| Ministral 3 14B | 34tok/sQ8_0 | 71tok/sQ8_0 |
| Qwen3 14B | 31tok/sQ8_0 | 65tok/sQ8_0 |
| DeepSeek R1 0528 Qwen3 8B | 54tok/sQ8_0 | 112tok/sQ8_0 |
| Qwen3 8B | 54tok/sQ8_0 | 112tok/sQ8_0 |
| Phi-4 14B | 31tok/sQ8_0 | 64tok/sQ8_0 |
| Qwen 3.5 4B | 111tok/sQ8_0 | 231tok/sQ8_0 |
| Ministral 3 8B | 52tok/sQ8_0 | 109tok/sQ8_0 |
| Granite 4.2 8B | 50tok/sQ8_0 | 104tok/sQ8_0 |
| Gemma 4 E4B | 63tok/sQ8_0 | 131tok/sQ8_0 |
| LFM2.5 8B-A1B | 218tok/sQ8_0 | 455tok/sQ8_0 |
| Llama 3.1 8B | 69tok/sQ6_K | 145tok/sQ6_K |
| Qwen2.5 0.5B | 843tok/sQ8_0 | 1759tok/sQ8_0 |
| Qwen3.5 0.8B | 572tok/sQ8_0 | 1193tok/sQ8_0 |
| Qwen3.5 2B | 342tok/sQ5_K_M | 715tok/sQ5_K_M |
| DeepSeek R1 Distill Qwen 1.5B | 250tok/sQ8_0 | 522tok/sQ8_0 |
| MiniCPM5 2B | 176tok/sQ8_0 | 366tok/sQ8_0 |
| LFM2.5 2.6B | 158tok/sQ8_0 | 329tok/sQ8_0 |
| LFM2.5 VL 3B | 158tok/sQ8_0 | 329tok/sQ8_0 |
| Spark X2.5 4B | 157tok/smlx-4bit | 327tok/smlx-4bit |
| granite vision 4.1 4b | 124tok/sQ8_0 | 259tok/sQ8_0 |
| granite 4.2 3b | 117tok/sQ8_0 | 243tok/sQ8_0 |
| LFM2.5 24B A2B | 114tok/sQ8_0 | 238tok/sQ8_0 |
| Mellum2.1 12B A2.5B Thinking | 113tok/sQ8_0 | 236tok/sQ8_0 |
| FrogNano 4B 2609 | 108tok/sQ8_0 | 225tok/sQ8_0 |
| Kimi VL A3B Thinking 2506 | 95tok/sQ8_0 | 199tok/sQ8_0 |
| Kimi VL A3B | 91tok/smlx-8bit | 190tok/smlx-8bit |
| DeepSeek R1 Distill Qwen 7B | 62tok/sQ8_0 | 130tok/sQ8_0 |
| GLM 4 9B 0414 | 62tok/smlx-5bit | 129tok/smlx-5bit |
| DeepSeek R1 Distill Llama 8B | 55tok/sQ8_0 | 116tok/sQ8_0 |
| Ministral 3 8B Reasoning 2512 | 52tok/sQ8_0 | 109tok/sQ8_0 |
| GLM 4.6V Flash | 52tok/sQ8_0 | 108tok/sQ8_0 |
| GLM 4.1V 9B Thinking | 44tok/smlx-8bit | 92tok/smlx-8bit |
| DeepSeek R1 Distill Qwen 14B | 31tok/sQ8_0 | 64tok/sQ8_0 |
| Qwen3.8 27B OBLITERATED | 25tok/smlx-4bit | 52tok/smlx-4bit |
| MiniMax M2 | 23tok/sQ8_0 | won't run |
| MiniMax M2.1 | 23tok/sQ8_0 | won't run |
| MiniMax M2.5 | 23tok/sQ8_0 | offload |
| Devstral Small 2507 | 20tok/sQ8_0 | 42tok/sQ8_0 |
| Magistral Small 2506 | 20tok/sQ8_0 | 42tok/sQ8_0 |
| Magistral Small 2507 | 20tok/sQ8_0 | 42tok/sQ8_0 |
| Magistral Small 2509 | 20tok/sQ8_0 | 42tok/sQ8_0 |
| GLM 4.5V | 19tok/sQ8_0 | 67tok/sQ4_K_M |
| GLM 4.6V | 19tok/sQ8_0 | 61tok/sQ4_K_M |
| DeepSeek V4 Flash 0731 | 18tok/sMXFP4 | won't run |
| DeepSeek V4 Flash Vision Exp | 17tok/sMXFP4 | won't run |
| granite 4.2 30b | 16tok/sQ8_0 | 33tok/sQ8_0 |
| MiniMax M3 | 16tok/sUD-Q6_K | won't run |
| GLM 5 | 14tok/sMXFP4 | won't run |
| DeepSeek R1 | 14tok/sQ4_K_M | won't run |
| DeepSeek R1 0528 | 14tok/sQ4_K_M | won't run |
| DeepSeek R1 Zero | 14tok/sQ4_K_M | won't run |
| DeepSeek V3 0324 | 14tok/sQ4_K_M | won't run |
| DeepSeek V3.1 | 14tok/sQ4_K_M | won't run |
| DeepSeek V3.1 Terminus | 14tok/sQ4_K_M | won't run |
| DeepSeek V3.2 | 13tok/sQ4_K_M | won't run |
| GLM 4.5 | 8tok/sQ8_0 | won't run |
| GLM 4.6 | 7tok/sQ8_0 | won't run |
| GLM 4.7 | 7tok/sQ8_0 | won't run |
| DeepSeek R1 Distill Llama 70B | 7tok/sQ8_0 | 14tok/sQ8_0 |
| Kimi Dev 72B | 7tok/sQ8_0 | 14tok/sQ8_0 |
| Devstral 2 123B Instruct 2512 | 4tok/sQ8_0 | 12tok/sQ5_K_M |
| Mistral Medium 3.5 128B | 4tok/sQ8_0 | 12tok/sQ5_K_M |
| GLM 5.1 | won't run | won't run |
| GLM 5.2 | won't run | won't run |
| GLM 5.3 | won't run | won't run |
| Kimi K2 | won't run | won't run |
| Kimi K2 Instruct 0905 | won't run | won't run |
| Kimi K2 Thinking | won't run | won't run |
| Kimi K2.5 | won't run | won't run |
Questions
- Which is better for local AI, the Mac Studio M3 Ultra 512GB or the NVIDIA RTX PRO 6000 Blackwell?
- Choose the Mac Studio M3 Ultra 512GB to run bigger models. Choose the NVIDIA RTX PRO 6000 Blackwell for speed. Choose the NVIDIA RTX PRO 6000 Blackwell to spend less.
- Which can run bigger models, the Mac Studio M3 Ultra 512GB or the NVIDIA RTX PRO 6000 Blackwell?
- At 8k context the Mac Studio M3 Ultra 512GB runs 93 of the models we track fully in memory and the NVIDIA RTX PRO 6000 Blackwell 75; only the Mac Studio M3 Ultra 512GB holds GLM 5.3 Flash.
- Which is faster for local AI, the Mac Studio M3 Ultra 512GB or the NVIDIA RTX PRO 6000 Blackwell?
- On the models both run fully in memory, the NVIDIA RTX PRO 6000 Blackwell is about 2.1× as fast as the Mac Studio M3 Ultra 512GB (estimated, median across models).
- What is the price difference between the Mac Studio M3 Ultra 512GB and the NVIDIA RTX PRO 6000 Blackwell?
- In the UK the Mac Studio M3 Ultra 512GB costs £17,495 at eBay (used) and the NVIDIA RTX PRO 6000 Blackwell £14,499 at eBay (used): the NVIDIA RTX PRO 6000 Blackwell is £2,996 cheaper.