AMD Instinct MI300X 192GBBudget pick
C55.1 tok/s decode
~$15,000 MSRP
Can it run?
Tight fit
Using Q4_K_M in vLLM
Fit status
Tight fit
Decode
35.5 tok/s
TTFT
5453 ms
Safe context
19K
Memory
109.4 GB / 128.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs with offload (needs ~0.4 GB host RAM) | 34.4 tok/s | 8192 ms | 32K |
| Chat | C | Runs well | 34.5 tok/s | 3059 ms | 10K |
| Coding | C | Tight fit | 35.5 tok/s | 5453 ms | 19K |
| RAG | C | Runs with offload (needs ~0.4 GB host RAM) | 34.4 tok/s | 10240 ms | 32K |
| Reasoning | C | Tight fit | 34.5 tok/s | 6628 ms |
How Devstral 2 123B Instruct (123B params) fits at each quantization level on Gaudi 3 128GB (128.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 48.0 GB | Low | D38 |
Q3_K_S | 3 | 60.3 GB | Low | D40 |
NVFP4 | 4 |
huggingface-cli download devstral-2-123bUpgrade options
~$15,000 MSRP
NVIDIA B200 180GBBest value~$30,000 MSRP
NVIDIA H200 141GBBiggest leap~$30,000 MSRP
| 19K |
68.9 GB |
| Medium |
| C41 |
Q4_K_M | 4 | 75.0 GB | Medium | C42 |
Q5_K_M | 5 | 88.6 GB | High | C44 |
Q6_KBest for your GPU | 6 | 100.9 GB | High | C45 |
Q8_0 | 8 | 131.6 GB | Very High | F0 |
F16 | 16 | 252.2 GB | Maximum | F0 |