Yes, just — 3.6 GB spare
DeepSeek-R1-Distill-Llama 70B at Q4_K_M fits your L40S entirely on the GPU at 8K context, at an estimated 13 tokens per second. Past 19K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully on GPU
8K context
Q4_K_M · 39.7 GB
MIT / Llama 3.3 Community
Released Jan 2025
The strongest of the R1 distills, and the one that most needs 48 GB or more.
What hardware do I need for DeepSeek-R1-Distill-Llama 70B? →
The VRAM budget
weights 39.7 GB
Weights 39.7 GB
KV cache @ 8K 2.50 GB
Runtime overhead 0.6 GB
Free 3.6 GB of 46.4 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 69.9 GB | 73.0 GB | — | ~1.2 | −0.1% ppl | 26.6 GB over |
| Q6_K | 53.9 GB | 57.0 GB | — | ~2.7 | −0.4% ppl | 10.6 GB over |
| Q5_K_M | 46.6 GB | 49.7 GB | — | ~5.8 | −0.8% ppl | 3.3 GB over |
| Q4_K_M | 39.7 GB | 42.8 GB | 19K | 13 | −1.9% ppl | Recommended |
| Q3_K_M | 32.1 GB | 35.2 GB | 43K | 16 | −5.4% ppl | Long context |
| Q2_K | 27.5 GB | 30.6 GB | 58K | 19 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.
How to run it
$ llama-server \
-hf deepseek-ai/DeepSeek-R1-Distill-Llama-70B:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 39.7 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03Only 3.6 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 19K context.