Can RTX 4090 run DeepSeek-R1-Distill-Qwen-14B?
Comfortable: EXL2 4.65bpw needs 9.8 GB at 4K, batch 1; headroom 14.3 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 | 128 | 9.8 |
Quant memory at 4K
| Quant | Total GB | Verdict | Weight sizing |
|---|---|---|---|
| GGUF Q4_K_M | 10.1 | Comfortable | Estimated weights |
| EXL2 4.65bpw | 9.8 | Comfortable | Estimated weights |
| AWQ INT4 | 8.5 | Comfortable | Estimated weights |
Context capacity · EXL2 4.65bpw
| Context | Total GB | Verdict |
|---|---|---|
| 4K | 9.8 | Comfortable |
| 16K | 12.2 | Comfortable |
| 32K | 15.5 | Comfortable |
| 128K | 35.3 | Does not fit |
Generation speed
Measured: 128 tok/s, EXL2 4.65bpw, ExLlamaV2
EXL2 4.65bpw: 128 tok/s · Measured