AMD Instinct MI250X 128GBBudget pick
C36.9 tok/s decode
~$15,000 MSRP
Can it run?
Runs with offload (needs ~0.7 GB host RAM)
Using Q4_K_M in vLLM
Fit status
Runs with offload (needs ~0.7 GB host RAM)
Decode
47.6 tok/s
TTFT
4068 ms
Safe context
16K
Memory
97.1 GB / 96.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | D | Very compromised (needs ~10.9 GB host RAM) | 44.0 tok/s | 6400 ms | 27K |
| Chat | C | Tight fit | 47.9 tok/s | 2207 ms | 9K |
| Coding | C | Runs with offload (needs ~0.7 GB host RAM) | 47.6 tok/s | 4068 ms | 16K |
| RAG | D | Very compromised (needs ~10.9 GB host RAM) | 44.0 tok/s | 7999 ms | 27K |
| Reasoning | C | Runs with offload (needs ~0.7 GB host RAM) | 47.6 tok/s | 4808 ms | 16K |
How Command A 111B (111B params) fits at each quantization level on NVIDIA GH200 96GB (96.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 43.3 GB | Low | D39 |
Q3_K_S | 3 | 54.4 GB | Low | C42 |
NVFP4 | 4 | 62.2 GB | Medium | C43 |
Q4_K_MBest for your GPU | 4 | 67.7 GB | Medium | C44 |
Q5_K_M | 5 | 79.9 GB | High | C44 |
Q6_K | 6 | 91.0 GB | High | C44 |
Q8_0 | 8 | 118.8 GB | Very High | F0 |
F16 | 16 | 227.6 GB | Maximum | F0 |
ollama run command-a-111bhuggingface-cli download command-a-111bUpgrade options
~$15,000 MSRP
~$15,000 MSRP
NVIDIA H200 141GBBiggest leap~$30,000 MSRP
NVIDIA H200 PCIe 141GBNVIDIA upgrade~$30,000 MSRP