RTX 5090 32GBBudget pickB93.7 tok/s decode
~$1,999 MSRP
Can it run?
Tight fit
Using Q4_K_M in Ollama
Fit status
Tight fit
Decode
52.4 tok/s
TTFT
3692 ms
Safe context
20K
Memory
19.7 GB / 24.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs with offload | 56.8 tok/s | 4957 ms | 33K |
| Chat | C | Runs well | 56.8 tok/s | 1859 ms | 11K |
| Coding | C | Tight fit | 52.4 tok/s | 3692 ms | 20K |
| RAG | C | Runs with offload | 56.8 tok/s | 6196 ms | 33K |
| Reasoning | C | Tight fit | 56.8 tok/s | 4027 ms | 20K |
How Codestral 21B Pruned i1 (21B params) fits at each quantization level on NVIDIA A30 24GB (24.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 8.2 GB | Low | D36 |
Q3_K_S | 3 | 10.3 GB | Low | D38 |
NVFP4 | 4 |
Upgrade options
RTX 5090 32GBBudget pick~$1,999 MSRP
NVIDIA A100 40GBBest value~$10,000 MSRP
RTX PRO 4500 Blackwell 32GB11.8 GB |
| Medium |
| D39 |
Q4_K_M | 4 | 12.8 GB | Medium | C40 |
Q5_K_M | 5 | 15.1 GB | High | C42 |
Q6_KBest for your GPU | 6 | 17.2 GB | High | C44 |
Q8_0 | 8 | 22.5 GB | Very High | C44 |
F16 | 16 | 43.1 GB | Maximum | F0 |