AMD Instinct MI250X 128GBBudget pick
C186 tok/s decode
~$15,000 MSRP
Can it run?
Runs well
Using Q4_K_M in vLLM
Fit status
Runs well
Decode
34.6 tok/s
TTFT
5599 ms
Safe context
45K
Memory
33.1 GB / 92.2 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 38.7 tok/s | 7267 ms | 81K |
| Chat | C | Runs well | 34.6 tok/s | 3054 ms | 24K |
| Coding | C | Runs well | 34.6 tok/s | 5599 ms | 45K |
| RAG | C | Runs well | 34.6 tok/s | 10181 ms | 81K |
| Reasoning | C | Runs well | 34.6 tok/s | 6617 ms | 45K |
How Codestral 2 25.08 (22B params) fits at each quantization level on Mac Studio M2 Ultra 128GB (92.2 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 8.6 GB | Low | D31 |
Q3_K_S | 3 | 10.8 GB | Low | D32 |
NVFP4 | 4 |
huggingface-cli download codestral-2-25.08Upgrade options
~$15,000 MSRP
NVIDIA H200 141GBBest value~$30,000 MSRP
NVIDIA GH200 96GBBiggest leap
12.3 GB |
| Medium |
| D32 |
Q4_K_M | 4 | 13.4 GB | Medium | D32 |
Q5_K_M | 5 | 15.8 GB | High | D33 |
Q6_K | 6 | 18.0 GB | High | D34 |
Q8_0 | 8 | 23.5 GB | Very High | D35 |
F16Best for your GPU | 16 | 45.1 GB | Maximum | D40 |