AMD Instinct MI350X 288GBBudget pick
C77.8 tok/s decode
~$8,000 MSRP
Can it run?
Runs well
Using Q4_K_M in vLLM
Fit status
Runs well
Decode
58.4 tok/s
TTFT
3316 ms
Safe context
34K
Memory
122.2 GB / 256.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 58.4 tok/s | 4824 ms | 58K |
| Chat | C | Runs well | 58.4 tok/s | 1809 ms | 18K |
| Coding | C | Runs well | 58.4 tok/s | 3316 ms | 34K |
| RAG | C | Runs well | 58.4 tok/s | 6030 ms | 58K |
| Reasoning | C | Runs well | 58.4 tok/s | 3919 ms | 34K |
How Devstral 2 123B Instruct (123B params) fits at each quantization level on AMD Instinct MI325X 256GB (256.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 48.0 GB | Low | D34 |
Q3_K_S | 3 | 60.3 GB | Low | D35 |
NVFP4 | 4 | 68.9 GB | Medium | D35 |
Q4_K_M | 4 | 75.0 GB | Medium | D36 |
Q5_K_M | 5 | 88.6 GB | High | D37 |
Q6_K | 6 | 100.9 GB | High | D38 |
Q8_0Best for your GPU | 8 | 131.6 GB | Very High | C41 |
F16 | 16 | 252.2 GB | Maximum | C45 |
huggingface-cli download devstral-2-123bUpgrade options
~$8,000 MSRP