Llama · Llama 3.1 · 405.9B parameters

Llama 3.1 405B VRAM requirements

Llama 3.1 405B has 126 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 229.5 GB.

Won't fit

234.4 GB of 21.8 GB · 1077%
0244 GB
Weights 229.5 GB
KV cache 3.9 GB
Runtime overhead 1.0 GB
Over the limit 212.7 GB

Short by 212.7 GB. You can run it with 9 of 126 layers on the RTX 4090 and the rest in system RAM, at roughly 0.18 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.

Generation0.18tok/s
Prompt processing85.4tok/s
Max context0tokens
KV per 1K tokens0GB

Every quantisation of Llama 3.1 405B on a RTX 4090

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

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 756.0 GB 760.9 GB Won't fit 0.05
INT8 / W8A8 8.50 400.7 GB 405.6 GB Won't fit 0.10
Q8_0 (GGUF) 8.50 400.7 GB 405.6 GB Won't fit 0.10
FP8 (E4M3) 8.00 378.0 GB 382.9 GB Won't fit 0.10
Q6_K 6.56 309.9 GB 314.9 GB Won't fit 0.13
Q5_K_M 5.67 267.9 GB 272.9 GB Won't fit 0.15
Q5_K_S 5.52 260.8 GB 265.8 GB Won't fit 0.15
Q4_K_M 4.85 229.5 GB 234.4 GB Won't fit 0.18
Q4_K_S 4.58 216.8 GB 221.8 GB Won't fit 0.19
Q4_0 4.55 215.4 GB 220.4 GB Won't fit 0.19
AWQ 4-bit 4.25 206.5 GB 211.5 GB Won't fit 0.20
GPTQ 4-bit 4.25 206.5 GB 211.5 GB Won't fit 0.20
MXFP4 4.25 206.5 GB 211.5 GB Won't fit 0.20
IQ4_XS 4.25 201.4 GB 206.4 GB Won't fit 0.20
Q3_K_M 3.91 185.5 GB 190.5 GB Won't fit 0.22
IQ3_M 3.70 175.7 GB 180.7 GB Won't fit 0.24
IQ3_XXS 3.06 145.8 GB 150.7 GB Won't fit 0.29
Q2_K 2.63 125.7 GB 130.6 GB Won't fit 0.34
IQ2_XXS 2.06 99.0 GB 104.0 GB Won't fit 0.45
IQ1_M 1.75 84.5 GB 89.5 GB Won't fit 0.54

Llama 3.1 405B on each GPU

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

GPUVRAMGB/sVerdictMax ctxtok/s
Mac Studio M3 Ultra 256GB 256 819 Won't fit 0.63
H100 SXM 80GB 80 3350 Won't fit 0.27
Mac Studio M4 Max 128GB 128 546 Won't fit 0.27
NVIDIA DGX Spark (GB10) 128 273 Won't fit 0.26
A100 80GB 80 2039 Won't fit 0.24
Ryzen AI Max+ 395 128GB 128 256 Won't fit 0.21
RTX A6000 48 768 Won't fit 0.20
RTX 5090 32 1792 Won't fit 0.20
Mac Mini M4 Pro 48GB 48 273 Won't fit 0.20
L40S 48 864 Won't fit 0.20
RTX 5080 16 960 Won't fit 0.19
RTX 5070 Ti 16 896 Won't fit 0.19
RTX 5060 Ti 16GB 16 448 Won't fit 0.19
RTX 5070 12 672 Won't fit 0.19
RTX 3090 24 936 Won't fit 0.18
RTX 4090 24 1008 Won't fit 0.18
RTX 3060 12GB 12 360 Won't fit 0.18
RTX 3080 10GB 10 760 Won't fit 0.17
RTX 4080 Super 16 736 Won't fit 0.17
RTX 4070 Ti Super 16 672 Won't fit 0.17
RTX 4060 Ti 16GB 16 288 Won't fit 0.17
RTX 4070 Super 12 504 Won't fit 0.17
RTX 4070 12 504 Won't fit 0.17
Radeon RX 7900 XTX 24 960 Won't fit 0.17
Arc B580 12 456 Won't fit 0.14

Architecture

Parameters405.9B
Layers126
Hidden size16384
Attention heads / KV heads128 / 8
Head dimension128
Vocabulary128,256
Trained context128K
KV cache per 1K tokens0 GB
Hugging Facemeta-llama/Llama-3.1-405B-Instruct

The Llama family

Meta's open-weight series, and the default target for most local tooling. Every Llama 3.x model uses grouped-query attention with 8 KV heads, so the cache stays modest even at 70B. Llama 4 moved to mixture-of-experts: Scout and Maverick occupy 109B and 400B of memory but read only 17B per token.

huggingface.co/meta-llama · llama.com · all 7 Llama models

Direct answers