Only with CPU offload
DeepSeek-R1-Distill-Qwen 32B at Q4_K_M needs 21.0 GB but only 18.4 GB is addressable, so about 14% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 9.5 tokens per second — usable for batch work, painful for chat.
86% on GPU
8K context
Q4_K_M · 18.4 GB
MIT
Released Jan 2025
MIT-licensed and close to the 70B distill on maths. A 24 GB card handles it at Q4.
What hardware do I need for DeepSeek-R1-Distill-Qwen 32B? →
Fits instead: DeepSeek-R1-Distill-Qwen 14B (10.4 GB) · DeepSeek-R1-Distill-Qwen 7B (5.3 GB)
The VRAM budget
weights 18.4 GB
Weights 18.4 GB
KV cache @ 8K 2.00 GB
Runtime overhead 0.6 GB
Over budget 2.6 GB past 18.4 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 32.5 GB | 35.1 GB | — | ~2.0 | −0.1% ppl | 16.7 GB over |
| Q6_K | 25.0 GB | 27.6 GB | — | ~3.5 | −0.4% ppl | 9.2 GB over |
| Q5_K_M | 21.7 GB | 24.3 GB | — | ~5.2 | −0.8% ppl | 5.9 GB over |
| Q4_K_M | 18.4 GB | 21.0 GB | — | ~9.5 | −1.9% ppl | 2.6 GB over |
| Q3_K_M | 14.9 GB | 17.5 GB | 11K | 32 | −5.4% ppl | Fits |
| Q2_K | 12.8 GB | 15.4 GB | 20K | 38 | −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-Qwen-32B:Q4_K_M \
-c 8192 -ngl 54
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 18.4 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.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.