RTX 5090 32GBBudget pickB72.9 tok/s decode
~$1,999 MSRP
Can it run?
Runs with offload (needs ~0.2 GB host RAM)
Using Q4_K_M in Ollama
Fit status
Runs with offload (needs ~0.2 GB host RAM)
Decode
39.4 tok/s
TTFT
4916 ms
Safe context
16K
Memory
24.3 GB / 24.0 GB
| Workload | Grade | Fit | Decode | TTFT | Context |
|---|---|---|---|---|---|
| Agentic Coding | D | Very compromised (needs ~2.6 GB host RAM) | 34.4 tok/s | 8178 ms | 27K |
| Chat | C | Tight fit | 39.8 tok/s | 2655 ms | 9K |
| Coding | C | Runs with offload (needs ~0.2 GB host RAM) | 39.4 tok/s | 4916 ms | 16K |
| RAG | D | Very compromised (needs ~2.6 GB host RAM) | 34.4 tok/s | 10222 ms | 27K |
| Reasoning | C | Runs with offload (needs ~0.2 GB host RAM) | 39.4 tok/s | 5810 ms | 16K |
How Qwen3.5 27B (27B params) fits at each quantization level on RTX 3090 24GB (24.0 GB usable).
| Quant | Bits | VRAM | Quality | Fit |
|---|---|---|---|---|
Q2_K | 2 | 10.5 GB | Low | D39 |
Q3_K_S | 3 | 13.2 GB | Low | C41 |
NVFP4 | 4 | 15.1 GB | Medium | C43 |
Q4_K_MBest for your GPU | 4 | 16.5 GB | Medium | C44 |
Q5_K_M | 5 | 19.4 GB | High | C45 |
Q6_K | 6 | 22.1 GB | High | C45 |
Q8_0 | 8 | 28.9 GB | Very High | F0 |
F16 | 16 | 55.4 GB | Maximum | F0 |
huggingface-cli download hf-unsloth--qwen3-5-27b-ggufUpgrade options
RTX 5090 32GBBudget pick~$1,999 MSRP
~$2,499 MSRP
NVIDIA A100 40GBBiggest leap~$10,000 MSRP
RTX PRO 4500 Blackwell 32GBNVIDIA upgrade