Mac Studio M2 Ultra 128GBBudget pick
C10.6 tok/s decode
~$3,999 MSRP
Can it run?
Runs with offload
Using Q4_K_M in Ollama
Fit status
Runs with offload
Decode
25.4 tok/s
TTFT
7634 ms
Safe context
16K
Memory
62.8 GB / 64.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | D | Very compromised (needs ~5.9 GB host RAM) | 22.4 tok/s | 12548 ms | 28K |
| Chat | C | Tight fit | 25.4 tok/s | 4164 ms | 9K |
| Coding | C | Runs with offload | 25.4 tok/s | 7634 ms | 16K |
| RAG | D | Very compromised (needs ~5.9 GB host RAM) | 22.4 tok/s | 15684 ms | 28K |
| Reasoning | C | Runs with offload | 25.4 tok/s | 9022 ms | 16K |
How Qwen 2.5 72B (72B params) fits at each quantization level on AMD Instinct MI210 64GB (64.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 28.1 GB | Low | D39 |
Q3_K_S | 3 | 35.3 GB | Low | C41 |
NVFP4 | 4 | 40.3 GB | Medium | C43 |
Q4_K_MBest for your GPU | 4 | 43.9 GB | Medium | C44 |
Q5_K_M | 5 | 51.8 GB | High | C45 |
Q6_K | 6 | 59.0 GB | High | C45 |
Q8_0 | 8 | 77.0 GB | Very High | F0 |
F16 | 16 | 147.6 GB | Maximum | F0 |
ollama run qwen-2.5-72bhuggingface-cli download qwen-2.5-72bUpgrade options
~$3,999 MSRP
~$3,999 MSRP
NVIDIA GH200 96GBBiggest leap