Can RTX 4090 run Qwen3 14B Instruct?
Comfortable: GGUF Q5_K_M needs 11.7 GB at 4K, batch 1; headroom 12.3 GB.
Text input, FP16 KV cache, one card. Unified-memory capacity uses the GPU guide’s usable limit; negative headroom may require host offload. Speed source settings may differ from this 4K estimate.
Measured runs
| Framework | Quant | tok/s | VRAM used (GB) |
|---|---|---|---|
| ExLlamaV2 | EXL2 4.65bpw | 118 | 10 |
Quant memory at 4K
| Quant | Total GB | Verdict | Weight sizing |
|---|---|---|---|
| GGUF Q4_K_M | 10.0 | Comfortable | Estimated weights |
| GGUF Q5_K_M | 11.7 | Comfortable | Estimated weights |
| AWQ INT4 | 8.4 | Comfortable | Estimated weights |
| EXL2 4.65bpw | 9.7 | Comfortable | Estimated weights |
Context capacity · GGUF Q5_K_M
| Context | Total GB | Verdict |
|---|---|---|
| 4K | 11.7 | Comfortable |
| 16K | 13.7 | Comfortable |
| 32K | 16.5 | Comfortable |
Generation speed
Measured: 118 tok/s, EXL2 4.65bpw, ExLlamaV2
GGUF Q5_K_M: 67–89 tok/s · Derived range