Can RTX 4090 run DeepSeek-R1-Distill-Llama-8B?
Comfortable: GGUF Q5_K_M needs 6.5 GB at 4K, batch 1; headroom 17.5 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) |
|---|---|---|---|
| llama.cpp | GGUF Q4_K_M | 145 | 5.6 |
Quant memory at 4K
| Quant | Total GB | Verdict | Weight sizing |
|---|---|---|---|
| GGUF Q4_K_M | 5.6 | Comfortable | Estimated weights |
| GGUF Q5_K_M | 6.5 | Comfortable | Estimated weights |
| EXL2 4.65bpw | 5.4 | Comfortable | Estimated weights |
| AWQ INT4 | 4.8 | Comfortable | Estimated weights |
Context capacity · GGUF Q5_K_M
| Context | Total GB | Verdict |
|---|---|---|
| 4K | 6.5 | Comfortable |
| 16K | 8.2 | Comfortable |
| 32K | 10.4 | Comfortable |
| 128K | 23.6 | Tight |
Generation speed
Measured: 145 tok/s, GGUF Q4_K_M, llama.cpp
GGUF Q5_K_M: 124–164 tok/s · Derived range