RTX 5090 32GBBudget pickB172.7 tok/s decode
~$1,999 MSRP
Can it run?
Runs with offload
Using Q4_K_M in vLLM
Fit status
Runs with offload
Decode
77.3 tok/s
TTFT
2504 ms
Safe context
16K
Memory
23.9 GB / 24.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | C | Runs with offload (needs ~0 GB host RAM) | 77.2 tok/s | 3647 ms | 32K |
| Chat | C | Runs with offload | 77.3 tok/s | 1366 ms | 8K |
| Coding | C | Runs with offload | 77.3 tok/s | 2504 ms | 16K |
| RAG | C | Runs with offload (needs ~0 GB host RAM) | 77.2 tok/s | 4559 ms | 32K |
| Reasoning | C | Runs with offload | 77.3 tok/s | 2960 ms | 16K |
How Qwen3-VL 30B A3B Instruct (30B params) fits at each quantization level on RTX A5000 24GB (24.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 11.7 GB | Low | D40 |
Q3_K_S | 3 | 14.7 GB | Low | C43 |
NVFP4Best for your GPU | 4 | 16.8 GB | Medium | C45 |
Q4_K_M | 4 | 18.3 GB | Medium | C45 |
Q5_K_M | 5 | 21.6 GB | High | C45 |
Q6_K | 6 | 24.6 GB | High | F0 |
Q8_0 | 8 | 32.1 GB | Very High | F0 |
F16 | 16 | 61.5 GB | Maximum | F0 |
huggingface-cli download qwen-3-vl-30b-a3bUpgrade options
RTX 5090 32GBBudget pick~$1,999 MSRP
~$2,499 MSRP
RTX PRO 4500 Blackwell 32GBNVIDIA upgrade