No — not on this device
DeepSeek-R1 671B at Q4_K_M needs 378.4 GB against 8.8 GB usable, and the shortfall of 369.6 GB is more than 128 GB of system RAM can cover at a tolerable speed. A smaller sibling or a lower quantisation is the honest answer here.
The January 2025 moment. Multi-head latent attention keeps its KV cache tiny; the weights do not.
What hardware do I need for DeepSeek-R1 671B? →
Fits instead: DeepSeek-R1-Distill-Qwen 7B (5.3 GB)
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 664.0 GB | 665.1 GB | — | ~0.4 | −0.1% ppl | 656.3 GB over |
| Q6_K | 512.4 GB | 513.6 GB | — | ~0.5 | −0.4% ppl | 504.8 GB over |
| Q5_K_M | 442.9 GB | 444.0 GB | — | ~0.6 | −0.8% ppl | 435.2 GB over |
| Q4_K_M | 377.3 GB | 378.4 GB | — | ~0.7 | −1.9% ppl | 369.6 GB over |
| Q3_K_M | 305.4 GB | 306.6 GB | — | ~0.8 | −5.4% ppl | 297.8 GB over |
| Q2_K | 261.7 GB | 262.8 GB | — | ~1.0 | −15% ppl | 254.0 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. This model uses multi-head latent attention, so its cache is a compressed latent rather than full K and V.
How to run it
$ llama-server \
-hf deepseek-ai/DeepSeek-R1:Q4_K_M \
-c 8192 -ngl 1
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.