Can it run?
Runs well
Using Q4_K_M in vLLM
Fit status
Runs well
Decode
58.7 tok/s
TTFT
3300 ms
Safe context
32K
Memory
24.1 GB / 48.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs well | 58.7 tok/s | 4801 ms | 56K |
| Chat | C | Runs well | 58.7 tok/s | 1800 ms | 17K |
| Coding | C | Runs well | 58.7 tok/s | 3300 ms | 32K |
| RAG | C | Runs well | 58.7 tok/s | 6001 ms | 56K |
| Reasoning | C | Runs well | 58.7 tok/s | 3901 ms | 32K |
How Codestral 2 25.08 (22B params) fits at each quantization level on RTX 6000 Ada 48GB (48.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 8.6 GB | Low | D33 |
Q3_K_S | 3 | 10.8 GB | Low | D34 |
NVFP4 | 4 | 12.3 GB | Medium | D35 |
Q4_K_M | 4 | 13.4 GB | Medium | D35 |
Q5_K_M | 5 | 15.8 GB | High | D36 |
Q6_K | 6 | 18.0 GB | High | D37 |
Q8_0Best for your GPU | 8 | 23.5 GB | Very High | D40 |
F16 | 16 | 45.1 GB | Maximum | C44 |
huggingface-cli download codestral-2-25.08