Can it run?
Runs well
Using Q4_K_M in vLLM
Fit status
Runs well
Decode
96.1 tok/s
TTFT
2014 ms
Safe context
23K
Memory
22.5 GB / 32.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | B | Runs well | 96.1 tok/s | 2929 ms | 40K |
| Chat | B | Runs well | 96.1 tok/s | 1098 ms | 12K |
| Coding | B | Runs well | 96.1 tok/s | 2014 ms | 23K |
| RAG | B | Runs well | 96.1 tok/s | 3661 ms | 40K |
| Reasoning | B | Runs well | 96.1 tok/s | 2380 ms | 23K |
How Codestral 2 25.08 (22B params) fits at each quantization level on RTX 5090 32GB (32.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 8.6 GB | Low | D35 |
Q3_K_S | 3 | 10.8 GB | Low | D37 |
NVFP4 | 4 |
huggingface-cli download codestral-2-25.08| Medium |
| D38 |
Q4_K_M | 4 | 13.4 GB | Medium | D38 |
Q5_K_M | 5 | 15.8 GB | High | D40 |
Q6_K | 6 | 18.0 GB | High | C41 |
Q8_0Best for your GPU | 8 | 23.5 GB | Very High | C44 |
F16 | 16 | 45.1 GB | Maximum | F0 |