Only with CPU offload
DeepSeek-R1-Distill-Llama 70B at Q4_K_M needs 42.8 GB but only 22.4 GB is addressable, so about 51% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 1.5 tokens per second — usable for batch work, painful for chat.
The strongest of the R1 distills, and the one that most needs 48 GB or more.
Hardware requirements What hardware do I need to run DeepSeek-R1-Distill-Llama 70B? →At Q4_K_M with an 8K context: a 64 GB GPU (Instinct MI210) or a 96 GB Mac (M2 Max). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.
Fits instead: DeepSeek-R1-Distill-Qwen 32B (21.0 GB) · DeepSeek-R1-Distill-Qwen 14B (10.4 GB)
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 69.9 GB | 73.0 GB | — | ~0.7 | −0.1% ppl | 50.6 GB over |
| Q6_K | 53.9 GB | 57.0 GB | — | ~0.9 | −0.4% ppl | 34.6 GB over |
| Q5_K_M | 46.6 GB | 49.7 GB | — | ~1.2 | −0.8% ppl | 27.3 GB over |
| Q4_K_M | 39.7 GB | 42.8 GB | — | ~1.5 | −1.9% ppl | 20.4 GB over |
| Q3_K_M | 32.1 GB | 35.2 GB | — | ~2.2 | −5.4% ppl | 12.8 GB over |
| Q2_K | 27.5 GB | 30.6 GB | — | ~3.0 | −15% ppl | 8.2 GB over |
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 38
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
Related models on an L4
| Model | Total | Tok/s est. | |
|---|---|---|---|
| DeepSeek-R1-Distill-Qwen 32B 32.8B · same family |
Q4_K_M · 21.0 GB | 9.8 | Tight fit |
| DeepSeek-R1-Distill-Qwen 14B 14.8B · same family |
Q4_K_M · 10.4 GB | 22 | Runs great |
| DeepSeek-R1-Distill-Qwen 7B 7.62B · same family |
Q4_K_M · 5.3 GB | 42 | Runs great |
| DeepSeek-R1 671B 671B MoE · same family |
Q4_K_M · 378.4 GB | — | Won't fit |
| Llama 3.3 70B Instruct 70.6B · similar size |
Q4_K_M · 42.8 GB | ~1.5 | CPU offload |
| Llama 4 Scout 109B-A17B 109B MoE · similar size |
Q4_K_M · 63.4 GB | — | Won't fit |
| gpt-oss 120B 117B MoE · similar size |
MXFP4 · 61.4 GB | — | Won't fit |
DeepSeek-R1-Distill-Llama 70B on other hardware
| Device | Total | Tok/s est. | |
|---|---|---|---|
| GeForce RTX 4060 Laptop 7.0 GB usable · 256 GB/s |
42.8 GB | — | Won't fit |
| GeForce RTX 3060 12 GB 10.6 GB usable · 360 GB/s |
42.8 GB | — | Won't fit |
| M4 · 16 GB 10.7 GB usable · 120 GB/s |
42.8 GB | — | Won't fit |
| GeForce RTX 4090 22.4 GB usable · 1008 GB/s |
42.8 GB | ~1.7 | CPU offload |
| Radeon RX 7900 XTX 22.4 GB usable · 960 GB/s |
42.8 GB | ~1.7 | CPU offload |
| Instinct MI210 62.4 GB usable · 1638 GB/s |
42.8 GB | 25 | Runs great |
| M3 Max · 64 GB 48.0 GB usable · 400 GB/s |
42.8 GB | 5.6 | Tight fit |
| M2 Max · 96 GB 72.0 GB usable · 400 GB/s |
42.8 GB | 5.6 | Runs great |