AMD Instinct MI300A 128GBBudget pick
C84.5 tok/s decode
Can it run?
Runs well
Using Q4_K_M in vLLM
Fit status
Runs well
Decode
26.2 tok/s
TTFT
7395 ms
Safe context
23K
Memory
67.2 GB / 96.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 26.2 tok/s | 10756 ms | 33K |
| Chat | C | Runs well | 26.2 tok/s | 4034 ms | 12K |
| Coding | C | Runs well | 26.2 tok/s | 7395 ms | 23K |
| RAG | C | Runs well | 26.2 tok/s | 13445 ms | 33K |
| Reasoning | C | Runs well | 26.2 tok/s | 8739 ms | 23K |
How Qwen 2.5 VL 72B (72B params) fits at each quantization level on RTX PRO 6000 Blackwell Server Edition 96GB (96.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 28.1 GB | Low | D36 |
Q3_K_S | 3 | 35.3 GB | Low | D37 |
NVFP4 | 4 |
ollama run qwen-2.5-vl-72bhuggingface-cli download qwen-2.5-vl-72bUpgrade options
40.3 GB |
| Medium |
| D38 |
Q4_K_M | 4 | 43.9 GB | Medium | D39 |
Q5_K_M | 5 | 51.8 GB | High | C41 |
Q6_K | 6 | 59.0 GB | High | C43 |
Q8_0Best for your GPU | 8 | 77.0 GB | Very High | C44 |
F16 | 16 | 147.6 GB | Maximum | F0 |