AMD Instinct MI350X 288GBBudget pick
C77.8 tok/s decode
~$8,000 MSRP
Can it run?
Very compromised (needs ~7.2 GB host RAM)
Using Q4_K_M in vLLM
Fit status
Very compromised (needs ~7.2 GB host RAM)
Decode
41.1 tok/s
TTFT
4710 ms
Safe context
14K
Memory
106.2 GB / 96.0 GB
Offload
10%
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | F | Too heavy | 43.2 tok/s | 6521 ms | 24K |
| Chat | C | Runs with offload (needs ~0.5 GB host RAM) | 43.0 tok/s | 2454 ms | 8K |
| Coding | C | Very compromised (needs ~7.2 GB host RAM) | 41.1 tok/s | 4710 ms | 14K |
| RAG | F | Too heavy | 43.2 tok/s | 8151 ms | 24K |
| Reasoning | C | Very compromised (needs ~7.2 GB host RAM) | 41.1 tok/s | 5567 ms | 14K |
How Devstral 2 123B Instruct (123B params) fits at each quantization level on NVIDIA GH200 96GB (96.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 48.0 GB | Low | C40 |
Q3_K_S | 3 | 60.3 GB | Low | C43 |
NVFP4 | 4 | 68.9 GB | Medium | C45 |
Q4_K_MBest for your GPU | 4 | 75.0 GB | Medium | C45 |
Q5_K_M | 5 | 88.6 GB | High | C45 |
Q6_K | 6 | 100.9 GB | High | F0 |
Q8_0 | 8 | 131.6 GB | Very High | F0 |
F16 | 16 | 252.2 GB | Maximum | F0 |
huggingface-cli download devstral-2-123bUpgrade options
~$8,000 MSRP
~$15,000 MSRP
NVIDIA B200 180GBBiggest leap~$30,000 MSRP
NVIDIA H200 141GBNVIDIA upgrade~$30,000 MSRP