RTX 5090 32GBBudget pickB89.5 tok/s decode
~$1,999 MSRP
Can it run?
Tight fit
Using Q4_K_M in vLLM
Fit status
Tight fit
Decode
24.5 tok/s
TTFT
7913 ms
Safe context
18K
Memory
21.7 GB / 24.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs with offload (needs ~0.6 GB host RAM) | 17.7 tok/s | 15902 ms | 31K |
| Chat | C | Tight fit | 18.3 tok/s | 5755 ms | 10K |
| Coding | C | Tight fit | 24.5 tok/s | 7913 ms | 18K |
| RAG | C | Runs with offload (needs ~0.6 GB host RAM) | 17.7 tok/s | 19878 ms | 31K |
| Reasoning | C | Tight fit | 18.3 tok/s | 12470 ms |
How Codestral 2 25.08 (22B params) fits at each quantization level on Intel Arc Pro B60 24GB (24.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 8.6 GB | Low | D37 |
Q3_K_S | 3 | 10.8 GB | Low | D39 |
NVFP4 | 4 |
huggingface-cli download codestral-2-25.08Upgrade options
RTX 5090 32GBBudget pick~$1,999 MSRP
~$2,499 MSRP
| 18K |
12.3 GB |
| Medium |
| C40 |
Q4_K_M | 4 | 13.4 GB | Medium | C41 |
Q5_K_M | 5 | 15.8 GB | High | C43 |
Q6_KBest for your GPU | 6 | 18.0 GB | High | C44 |
Q8_0 | 8 | 23.5 GB | Very High | C44 |
F16 | 16 | 45.1 GB | Maximum | F0 |