AMD Instinct MI250X 128GBBudget pick
C38.8 tok/s decode
~$15,000 MSRP
Can it run?
Tight fit
Using Q3_K_M in Ollama
Fit status
Tight fit
Decode
50.4 tok/s
TTFT
3841 ms
Safe context
17K
Memory
89.6 GB / 96.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Very compromised (needs ~7 GB host RAM) | 47.5 tok/s | 5934 ms | 28K |
| Chat | C | Tight fit | 50.4 tok/s | 2095 ms | 10K |
| Coding | C | Tight fit | 50.4 tok/s | 3841 ms | 17K |
| RAG | C | Very compromised (needs ~7 GB host RAM) | 47.5 tok/s | 7417 ms | 28K |
| Reasoning | C | Tight fit | 50.4 tok/s | 4539 ms | 17K |
How Qwen3.5 122B A10B (122B params) fits at each quantization level on NVIDIA GH200 96GB (96.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 47.6 GB | Low | C40 |
Q3_K_S | 3 | 59.8 GB | Low | C43 |
NVFP4 | 4 |
huggingface-cli download hf-unsloth--qwen3-5-122b-a10b-ggufUpgrade options
~$15,000 MSRP
NVIDIA H200 141GBBest value~$30,000 MSRP
NVIDIA H200 PCIe 141GBNVIDIA upgrade~$30,000 MSRP
68.3 GB |
| Medium |
| C45 |
Q4_K_MBest for your GPU | 4 | 74.4 GB | Medium | C45 |
Q5_K_M | 5 | 87.8 GB | High | C45 |
Q6_K | 6 | 100.0 GB | High | F0 |
Q8_0 | 8 | 130.5 GB | Very High | F0 |
F16 | 16 | 250.1 GB | Maximum | F0 |