CanItRun Logocanitrun.

Llama 4 Scout 109B

Llama 4 Scout 109B needs roughly 77.3 GB VRAM at Q4_K_M quantization (247.2 GB at FP16). 29 GPUs we track can run it fully in VRAM at 8k context.

29 GPUs run this natively · 10 with CPU offload

Meta109B params17B active (MoE)9766k contextLlama 4 CommunityCommercial use ok

Llama 4 Scout 109B is a Mixture of Experts (MoE) model with 109B total parameters but only 17B active per token developed by Meta. April 2025 MoE model with 109B total parameters but only 17B active per token. 10M context window is industry-leading.

To run Llama 4 Scout 109B locally: Q4 needs ~60-70GB VRAM — requires 80GB GPU or Mac Studio. The 10M context is theoretical; KV cache limits practical usage. As a MoE model, inference speed depends on active parameters (17B) rather than total size.

MMLU-Pro 70.0% with MoE efficiency — combines frontier quality with practical inference costs.

VRAM at each quantization

Numbers here are computed at 8k context. Because KV cache grows linearly with context length, expect higher totals at longer sequence lengths.

QuantWeightsKV cacheTotal
FP32436.0 GB2.68 GB491.3 GB
BF16218.0 GB2.68 GB247.2 GB
FP16218.0 GB2.68 GB247.2 GB
Q8_0115.9 GB2.68 GB132.8 GB
Q6_K89.5 GB2.68 GB103.2 GB
Q5_K_M77.6 GB2.68 GB89.9 GB
Q4_K_Mrec66.4 GB2.68 GB77.3 GB
Q3_K_M52.4 GB2.68 GB61.7 GB
Q2_K41.5 GB2.68 GB49.5 GB
NVFP4cuda54.5 GB2.68 GB64.0 GB

KV cache is calculated at 8k context (FP16). Note that NVFP4 only runs on CUDA GPUs. Turn on TurboQuant in the calculator above for lower KV cache estimates.

Benchmarks

MMLU-Pro
74.3

GPUs that run Llama 4 Scout 109B natively (29)

Plus 10 GPUs that run it with CPU offload (slower)

Notes

16 experts, 2 active. 10M context; KV cache limits practical context to much less.

Hugging Face ↗Ollama ↗Released 2025-04-05

Compare Llama 4 Scout 109B with other models

Continue reading

vram-guides10 min
Llama Family Guide: Llama 3 to Llama 4 — VRAM & Hardware
hardware10 min
Best GPUs $1000+ for LLMs (2026)
hardware11 min
Apple Silicon for LLMs: M1-M5 Complete Guide

Frequently asked questions

What are the VRAM requirements for Llama 4 Scout 109B?
Llama 4 Scout 109B requires approximately 77.4 GB of VRAM at Q4_K_M quantization, 132.8 GB at Q8, and 247.2 GB at FP16. These numbers assume 8k context window; VRAM scales linearly with context length due to the KV cache.
How many parameters does Llama 4 Scout 109B have?
Llama 4 Scout 109B has 109 billion total parameters, but only 17 billion are active per token thanks to its Mixture of Experts (MoE) architecture. This makes inference significantly faster than the total parameter count suggests.
How capable is Llama 4 Scout 109B?
Llama 4 Scout 109B achieves an MMLU-Pro score of 74.3, placing it among the most capable open-weight models available — competitive with frontier systems on general knowledge and reasoning.
Can Llama 4 Scout 109B run on a 16 GB GPU?
No. At Q4_K_M, Llama 4 Scout 109B needs 77.4 GB of VRAM — more than 16 GB. You will need a multi-GPU server.
Can Llama 4 Scout 109B run on a 24 GB GPU?
No. Even at Q4_K_M, Llama 4 Scout 109B needs 77.4 GB. Consider a multi-GPU server with 80 GB+ total VRAM.
What is the smallest quantization for Llama 4 Scout 109B that fits in 24 GB of VRAM?
Llama 4 Scout 109B cannot fit in 24 GB of VRAM at any standard quantization level. The minimum needed is 49.5 GB at Q2_K.
What GPU do I need to run Llama 4 Scout 109B locally?
You need a multi-GPU server. At Q4_K_M, Llama 4 Scout 109B needs 77.4 GB VRAM, more than any single consumer GPU. Consider 2–4× H100 or A100 GPUs.