Llama · 109B parameters · 17B active

Llama 4 Scout 109B-A17B VRAM requirements

Llama 4 Scout 109B-A17B has 48 layers and uses grouped-query attention (8 KV heads). At Q4_K_M the weights come to 61.7 GB.

Won't fit

64.0 GB of 21.8 GB · 294%
067 GB
Weights 61.7 GB
KV cache 1.5 GB
Runtime overhead 0.8 GB
Over the limit 42.3 GB

Short by 42.3 GB. You can run it with 15 of 48 layers on the RTX 4090 and the rest in system RAM, at roughly 5.18 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.

Generation5.18tok/s
Prompt processing2038tok/s
Max context0tokens
KV per 1K tokens0GB

Every quantisation of Llama 4 Scout 109B-A17B on a RTX 4090

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

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 203.0 GB 205.4 GB Won't fit — 1.28
INT8 / W8A8 8.50 107.4 GB 109.7 GB Won't fit — 2.58
Q8_0 (GGUF) 8.50 107.4 GB 109.7 GB Won't fit — 2.58
FP8 (E4M3) 8.00 101.5 GB 103.9 GB Won't fit — 2.79
Q6_K 6.56 83.2 GB 85.6 GB Won't fit — 3.53
Q5_K_M 5.67 71.9 GB 74.3 GB Won't fit — 4.15
Q5_K_S 5.52 70.0 GB 72.4 GB Won't fit — 4.37
Q4_K_M 4.85 61.7 GB 64.0 GB Won't fit — 5.18
Q4_K_S 4.58 58.3 GB 60.7 GB Won't fit — 5.45
Q4_0 4.55 58.0 GB 60.3 GB Won't fit — 5.64
AWQ 4-bit 4.25 56.8 GB 59.1 GB Won't fit — 5.75
GPTQ 4-bit 4.25 56.8 GB 59.1 GB Won't fit — 5.75
MXFP4 4.25 56.8 GB 59.1 GB Won't fit — 5.75
IQ4_XS 4.25 54.2 GB 56.6 GB Won't fit — 6.17
Q3_K_M 3.91 50.0 GB 52.3 GB Won't fit — 6.84
IQ3_M 3.70 47.4 GB 49.7 GB Won't fit — 7.40
IQ3_XXS 3.06 39.4 GB 41.8 GB Won't fit — 9.93
Q2_K 2.63 34.1 GB 36.4 GB Won't fit — 13.1
IQ2_XXS 2.06 27.0 GB 29.3 GB Won't fit — 22.2
IQ1_M 1.75 23.1 GB 25.4 GB Won't fit — 37.8

Llama 4 Scout 109B-A17B 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 63K 177
A100 80GB 80 2039 Runs comfortably 63K 100
Mac Studio M3 Ultra 256GB 256 819 Runs comfortably 692K 38.5
Mac Studio M4 Max 128GB 128 546 Runs comfortably 180K 28.7
NVIDIA DGX Spark (GB10) 128 273 Runs comfortably 180K 14.8
Ryzen AI Max+ 395 128GB 128 256 Runs comfortably 180K 11.1
RTX A6000 48 768 Won't fit — 9.64
L40S 48 864 Won't fit — 9.40
RTX 5090 32 1792 Won't fit — 6.75
Mac Mini M4 Pro 48GB 48 273 Won't fit — 6.52
RTX 3090 24 936 Won't fit — 5.39
RTX 4090 24 1008 Won't fit — 5.18
Radeon RX 7900 XTX 24 960 Won't fit — 4.91
RTX 5080 16 960 Won't fit — 4.91
RTX 5070 Ti 16 896 Won't fit — 4.90
RTX 5060 Ti 16GB 16 448 Won't fit — 4.81
RTX 5070 12 672 Won't fit — 4.57
RTX 4080 Super 16 736 Won't fit — 4.43
RTX 4070 Ti Super 16 672 Won't fit — 4.42
RTX 4060 Ti 16GB 16 288 Won't fit — 4.28
RTX 3060 12GB 12 360 Won't fit — 4.27
RTX 3080 10GB 10 760 Won't fit — 4.16
RTX 4070 Super 12 504 Won't fit — 4.12
RTX 4070 12 504 Won't fit — 4.12
Arc B580 12 456 Won't fit — 3.48

Architecture

Parameters109B
Active per token17B of 16 experts, top-1
Layers48
Hidden size5120
Attention heads / KV heads40 / 8
Head dimension128
Vocabulary202,048
Trained context10240K
KV cache per 1K tokens0 GB
Hugging Facemeta-llama/Llama-4-Scout-17B-16E-Instruct

Direct answers

Llama 4 Scout 109B-A17B on RTX 5090
See the verdictLlama 4 Scout 109B-A17B on RTX 4090
See the verdictLlama 4 Scout 109B-A17B on RTX 3090
See the verdictLlama 4 Scout 109B-A17B on RTX 5080
See the verdictLlama 4 Scout 109B-A17B on RTX 5070 Ti
See the verdictLlama 4 Scout 109B-A17B on RTX 5070
See the verdictLlama 4 Scout 109B-A17B on RTX 4070 Ti Super
See the verdictLlama 4 Scout 109B-A17B on RTX 4070 Super
See the verdictLlama 4 Scout 109B-A17B on RTX 4070
See the verdict