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
7.4 tok/s
TTFT
26082 ms
Safe context
24K
Memory
124.3 GB / 184.3 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 7.4 tok/s | 37938 ms | 41K |
| Chat | C | Runs well | 7.4 tok/s | 14227 ms | 13K |
| Coding | C | Runs well | 7.4 tok/s | 26082 ms | 24K |
| RAG | C | Runs well | 7.4 tok/s | 47423 ms | 41K |
| Reasoning | C | Runs well | 7.4 tok/s | 30825 ms | 24K |
How Devstral 2 123B Instruct (123B params) fits at each quantization level on Mac Studio M3 Ultra 256GB (184.3 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 48.0 GB | Low | D35 |
Q3_K_S | 3 | 60.3 GB | Low | D37 |
NVFP4 | 4 | 68.9 GB | Medium | D38 |
Q4_K_M | 4 | 75.0 GB | Medium | D38 |
Q5_K_M | 5 | 88.6 GB | High | D40 |
Q6_K | 6 | 100.9 GB | High | C41 |
Q8_0Best for your GPU | 8 | 131.6 GB | Very High | C45 |
F16 | 16 | 252.2 GB | Maximum | F0 |
huggingface-cli download devstral-2-123bUpgrade options
~$8,000 MSRP
~$15,000 MSRP
H100 NVL 188GBBiggest leap