DeepSeek · DeepSeek-R1 · 32.8B parameters

DeepSeek-R1-Distill-Qwen 32B VRAM requirements

DeepSeek-R1-Distill-Qwen 32B has 64 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 18.6 GB, and the best quantisation that fits a 24 GB card is Q4_K_M.

Model

Hardware

Quantisation

Context 8K

KV cache precision

Runtime

Batch

GPUs

Monitor plugged into this GPU

Fits, but tight

21.5 GB of 21.8 GB · 99%

022 GB

Weights 18.6 GB

KV cache 2.0 GB

Runtime overhead 0.8 GB

This fits with almost nothing to spare. A background application claiming VRAM will push it over. Drop to the next quantisation down, or quantise the KV cache to Q8_0 — that halves the cache for no meaningful quality loss.

Generation30.5tok/s

Prompt processing1058tok/s

Max context9Ktokens

KV per 1K tokens0GB

Every quantisation of DeepSeek-R1-Distill-Qwen 32B on a RTX 4090

Highlighted row is the highest quality that still fits at 8K context.

Quantisation	bpw	Weights	Total	Verdict	Max ctx	tok/s
FP16 / BF16	16.00	61.0 GB	63.9 GB	Won't fit	—	0.85
INT8 / W8A8	8.50	32.1 GB	34.9 GB	Won't fit	—	2.52
Q8_0 (GGUF)	8.50	32.1 GB	34.9 GB	Won't fit	—	2.52
FP8 (E4M3)	8.00	30.5 GB	33.3 GB	Won't fit	—	2.83
Q6_K	6.56	25.0 GB	27.9 GB	Won't fit	—	4.94
Q5_K_M	5.67	21.6 GB	24.5 GB	Won't fit	—	9.34
Q5_K_S	5.52	21.1 GB	23.9 GB	Won't fit	—	10.4
Q4_K_M	4.85	18.6 GB	21.5 GB	Fits, but tight	9K	30.5
AWQ 4-bit	4.25	18.3 GB	21.2 GB	Fits, but tight	10K	30.9
GPTQ 4-bit	4.25	18.3 GB	21.2 GB	Fits, but tight	10K	30.9
MXFP4	4.25	18.3 GB	21.2 GB	Fits, but tight	10K	30.9
Q4_K_S	4.58	17.6 GB	20.5 GB	Fits, but tight	13K	32.0
Q4_0	4.55	17.5 GB	20.4 GB	Fits, but tight	14K	32.2
IQ4_XS	4.25	16.4 GB	19.3 GB	Runs comfortably	18K	34.2
Q3_K_M	3.91	15.2 GB	18.0 GB	Runs comfortably	23K	36.8
IQ3_M	3.70	14.4 GB	17.3 GB	Runs comfortably	26K	38.6
IQ3_XXS	3.06	12.1 GB	15.0 GB	Runs comfortably	35K	45.3
Q2_K	2.63	10.6 GB	13.4 GB	Runs comfortably	41K	51.3
IQ2_XXS	2.06	8.5 GB	11.3 GB	Runs comfortably	50K	62.2
IQ1_M	1.75	7.4 GB	10.2 GB	Runs comfortably	54K	70.4

DeepSeek-R1-Distill-Qwen 32B on each GPU

Q4_K_M weights at 8K context, single card, monitor attached.

GPU	VRAM	GB/s	Verdict	Max ctx	tok/s
H100 SXM 80GB	80	3350	Runs comfortably	128K	114
A100 80GB	80	2039	Runs comfortably	128K	63.5
RTX 5090	32	1792	Runs comfortably	39K	59.0
RTX 4090	24	1008	Fits, but tight	9K	30.5
RTX 3090	24	936	Fits, but tight	9K	29.5
Radeon RX 7900 XTX	24	960	Fits, but tight	9K	27.6
L40S	48	864	Runs comfortably	99K	26.1
RTX A6000	48	768	Runs comfortably	99K	24.3
Mac Studio M3 Ultra 256GB	256	819	Runs comfortably	128K	23.9
Mac Studio M4 Max 128GB	128	546	Runs comfortably	128K	17.8
NVIDIA DGX Spark (GB10)	128	273	Runs comfortably	128K	9.17
Mac Mini M4 Pro 48GB	48	273	Runs comfortably	67K	8.92
Ryzen AI Max+ 395 128GB	128	256	Runs comfortably	128K	6.84
RTX 5080	16	960	Won't fit	—	4.98
RTX 5070 Ti	16	896	Won't fit	—	4.95
RTX 5060 Ti 16GB	16	448	Won't fit	—	4.50
RTX 4080 Super	16	736	Won't fit	—	4.39
RTX 4070 Ti Super	16	672	Won't fit	—	4.34
RTX 4060 Ti 16GB	16	288	Won't fit	—	3.71
RTX 5070	12	672	Won't fit	—	3.40
RTX 3060 12GB	12	360	Won't fit	—	3.06
RTX 4070 Super	12	504	Won't fit	—	3.02
RTX 4070	12	504	Won't fit	—	3.02
RTX 3080 10GB	10	760	Won't fit	—	2.79
Arc B580	12	456	Won't fit	—	2.54

Architecture

Parameters	32.8B
Layers	64
Hidden size	5120
Attention heads / KV heads	40 / 8
Head dimension	128
Vocabulary	152,064
Trained context	128K
KV cache per 1K tokens	0 GB
Hugging Face	deepseek-ai/DeepSeek-R1-Distill-Qwen-32B

The DeepSeek family

V3 and R1 share one 671B mixture-of-experts architecture with 37B active per token, and both use multi-head latent attention — a single compressed KV vector per layer instead of a full set of heads. That gives R1 roughly a fifth of the cache per token of a dense 70B. The R1 distills are ordinary Qwen and Llama models fine-tuned on R1 output, and they fit on one card.

huggingface.co/deepseek-ai · deepseek.com · all 5 DeepSeek models

DeepSeek-R1 671B-A37B
671B · 37B active · 61 layers DeepSeek-R1-Distill-Llama 8B
8.0B · 32 layers DeepSeek-V3 671B-A37B
671B · 37B active · 61 layers DeepSeek-Coder-V2-Lite 16B-A2.4B
15.7B · 2.4B active · 27 layers

Direct answers