C127.6 tok/s decode
~$15,000 MSRP
Can it run?
Runs well
Using Q4_K_M in vLLM
Fit status
Runs well
Decode
32.2 tok/s
TTFT
6014 ms
Safe context
40K
Memory
25.7 GB / 64.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 34.9 tok/s | 8075 ms | 70K |
| Chat | C | Runs well | 34.9 tok/s | 3028 ms | 21K |
| Coding | C | Runs well | 32.2 tok/s | 6014 ms | 40K |
| RAG | C | Runs well | 34.9 tok/s | 10094 ms | 70K |
| Reasoning | C | Runs well | 34.9 tok/s | 6561 ms | 40K |
How Codestral 2 25.08 (22B params) fits at each quantization level on NVIDIA A16 64GB (64.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 8.6 GB | Low | D32 |
Q3_K_S | 3 | 10.8 GB | Low | D33 |
NVFP4 | 4 |
huggingface-cli download codestral-2-25.08Upgrade options
~$15,000 MSRP
~$30,000 MSRP
~$40,000 MSRP
| Medium |
| D33 |
Q4_K_M | 4 | 13.4 GB | Medium | D34 |
Q5_K_M | 5 | 15.8 GB | High | D35 |
Q6_K | 6 | 18.0 GB | High | D35 |
Q8_0 | 8 | 23.5 GB | Very High | D37 |
F16Best for your GPU | 16 | 45.1 GB | Maximum | C44 |