AMD Instinct MI350X 288GBBudget pick
C77.8 tok/s decode
~$8,000 MSRP
Can it run?
Very compromised
Using Q4_K_M in vLLM
Fit status
Very compromised
Decode
4.5 tok/s
TTFT
43071 ms
Safe context
13K
Memory
110.5 GB / 92.2 GB
Offload
20%
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | F | Too heavy | 4.7 tok/s | 59534 ms | 23K |
| Chat | D | Very compromised | 4.6 tok/s | 22918 ms | 7K |
| Coding | D | Very compromised | 4.5 tok/s | 43071 ms | 13K |
| RAG | F | Too heavy | 4.7 tok/s | 74417 ms | 23K |
| Reasoning | D | Very compromised | 4.5 tok/s | 50902 ms | 13K |
How Devstral 2 123B Instruct (123B params) fits at each quantization level on MacBook Pro M4 Max 128GB (92.2 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 48.0 GB | Low | C41 |
Q3_K_S | 3 | 60.3 GB | Low | C44 |
NVFP4Best for your GPU |
huggingface-cli download devstral-2-123bUpgrade options
~$8,000 MSRP
~$15,000 MSRP
NVIDIA B200 180GBBiggest leap| 4 |
68.9 GB |
| Medium |
| C45 |
Q4_K_M | 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 |
~$30,000 MSRP