Qwen · Qwen2.5 · 72.7B parameters

Qwen2.5 72B VRAM requirements

Qwen2.5 72B has 80 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 41.2 GB.

Won't fit

44.6 GB of 21.8 GB · 205%
046 GB
Weights 41.2 GB
KV cache 2.5 GB
Runtime overhead 0.9 GB
Over the limit 22.9 GB

Short by 22.9 GB. You can run it with 35 of 80 layers on the RTX 4090 and the rest in system RAM, at roughly 1.53 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.

Generation1.53tok/s
Prompt processing477tok/s
Max context0tokens
KV per 1K tokens0GB

Every quantisation of Qwen2.5 72B on a RTX 4090

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

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 135.4 GB 138.8 GB Won't fit — 0.32
INT8 / W8A8 8.50 71.4 GB 74.8 GB Won't fit — 0.69
Q8_0 (GGUF) 8.50 71.4 GB 74.8 GB Won't fit — 0.69
FP8 (E4M3) 8.00 67.7 GB 71.1 GB Won't fit — 0.74
Q6_K 6.56 55.5 GB 58.9 GB Won't fit — 0.97
Q5_K_M 5.67 48.0 GB 51.4 GB Won't fit — 1.20
Q5_K_S 5.52 46.7 GB 50.1 GB Won't fit — 1.25
Q4_K_M 4.85 41.2 GB 44.6 GB Won't fit — 1.53
AWQ 4-bit 4.25 39.4 GB 42.8 GB Won't fit — 1.66
GPTQ 4-bit 4.25 39.4 GB 42.8 GB Won't fit — 1.66
MXFP4 4.25 39.4 GB 42.8 GB Won't fit — 1.66
Q4_K_S 4.58 39.0 GB 42.4 GB Won't fit — 1.68
Q4_0 4.55 38.8 GB 42.2 GB Won't fit — 1.69
IQ4_XS 4.25 36.3 GB 39.7 GB Won't fit — 1.91
Q3_K_M 3.91 33.6 GB 36.9 GB Won't fit — 2.21
IQ3_M 3.70 31.8 GB 35.2 GB Won't fit — 2.50
IQ3_XXS 3.06 26.6 GB 30.0 GB Won't fit — 3.85
Q2_K 2.63 23.1 GB 26.5 GB Won't fit — 5.98
IQ2_XXS 2.06 18.4 GB 21.8 GB Won't fit 8K 25.6
IQ1_M 1.75 15.9 GB 19.3 GB Runs comfortably 16K 34.7

Qwen2.5 72B on each GPU

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

GPUVRAMGB/sVerdictMax ctxtok/s
H100 SXM 80GB 80 3350 Runs comfortably 103K 53.8
A100 80GB 80 2039 Runs comfortably 103K 29.6
Mac Studio M3 Ultra 256GB 256 819 Runs comfortably 128K 11.1
L40S 48 864 Won't fit 7K 10.5
RTX A6000 48 768 Won't fit 7K 9.90
Mac Studio M4 Max 128GB 128 546 Runs comfortably 128K 8.24
NVIDIA DGX Spark (GB10) 128 273 Runs comfortably 128K 4.24
Ryzen AI Max+ 395 128GB 128 256 Runs comfortably 128K 3.16
RTX 5090 32 1792 Won't fit — 2.50
Mac Mini M4 Pro 48GB 48 273 Won't fit — 2.43
RTX 3090 24 936 Won't fit — 1.59
RTX 4090 24 1008 Won't fit — 1.53
Radeon RX 7900 XTX 24 960 Won't fit — 1.44
RTX 5080 16 960 Won't fit — 1.31
RTX 5070 Ti 16 896 Won't fit — 1.31
RTX 5060 Ti 16GB 16 448 Won't fit — 1.28
RTX 4080 Super 16 736 Won't fit — 1.18
RTX 4070 Ti Super 16 672 Won't fit — 1.18
RTX 5070 12 672 Won't fit — 1.16
RTX 4060 Ti 16GB 16 288 Won't fit — 1.13
RTX 3060 12GB 12 360 Won't fit — 1.08
RTX 3080 10GB 10 760 Won't fit — 1.06
RTX 4070 Super 12 504 Won't fit — 1.05
RTX 4070 12 504 Won't fit — 1.05
Arc B580 12 456 Won't fit — 0.89

Architecture

Parameters72.7B
Layers80
Hidden size8192
Attention heads / KV heads64 / 8
Head dimension128
Vocabulary152,064
Trained context128K
KV cache per 1K tokens0 GB
Hugging FaceQwen/Qwen2.5-72B-Instruct

The Qwen family

Alibaba's series, and the broadest size ladder available — Qwen3 runs from 0.6B to 32B dense, plus 30B-A3B and 235B-A22B as mixture-of-experts. The 152k vocabulary makes the embedding table a large share of a small model's file. Qwen2.5-Coder is the same architecture trained for code.

huggingface.co/Qwen · qwenlm.github.io · all 14 Qwen models

Qwen2.5 7B
7.6B · 28 layersQwen2.5 14B
14.8B · 48 layersQwen2.5 32B
32.8B · 64 layersQwen3 0.6B
600M · 28 layersQwen3 1.7B
1.7B · 28 layersQwen3 4B
4.0B · 36 layersQwen3 8B
8.2B · 36 layersQwen3 14B
14.8B · 40 layers

Direct answers

Qwen2.5 72B on RTX 5090
See the verdictQwen2.5 72B on RTX 4090
See the verdictQwen2.5 72B on RTX 3090
See the verdictQwen2.5 72B on RTX 5080
See the verdictQwen2.5 72B on RTX 5070 Ti
See the verdictQwen2.5 72B on RTX 5070
See the verdictQwen2.5 72B on RTX 4070 Ti Super
See the verdictQwen2.5 72B on RTX 4070 Super
See the verdictQwen2.5 72B on RTX 4070
See the verdict