RTX 4090 for local LLMs
4 measured runs. 24 GB of VRAM; memory measured at 962 GB/s, which sets the roofline for every model below.
| Model | Quant | Engine | tok/s | Roofline | Share |
|---|---|---|---|---|---|
| Qwen3-14B | AWQ | vLLM | 97.7 | 113.0 | 86% |
| Qwen3-14B | Q4_K_M | llama.cpp | 94.4 | 111.2 | 85% |
| Qwen3.8-27B | W4A16-AWQ | vLLM | 54.4 | 61.7 | 88% |
| Qwen3.8-27B | UD-Q4_K_M | llama.cpp | 50.4 | 61.1 | 82% |
Share = measured tok/s over the roofline: the bytes a token reads over this GPU's measured memory speed. Open a quant for the test machine and the per-role and per-kernel breakdown.