Llama 4 Scout Setup Guide: Run Meta's 109B MoE Model Locally

Llama 4 Scout is Meta's 109B parameter Mixture-of-Experts model with only 17B active parameters per forward pass. That means it runs surprisingly fast on consumer hardware. This guide covers every path from a single RTX 4090 to a Mac Studio Ultra.

Key insight: MoE models activate only a fraction of parameters per token. Llama 4 Scout activates ~17B of 109B parameters, making its VRAM footprint similar to a 17B dense model while retaining 109B knowledge. For hardware comparisons, see LLMHardware.io's GPU benchmark tool.

Hardware Requirements

Quantization	Model Size	Min Hardware	Speed
Q2_K	~35 GB	2x RTX 4090 or Mac Ultra	~18 tok/s
Q4_K_M	~55 GB	Mac Ultra M4 or 3x RTX 4090	~14 tok/s ★
Q5_K_M	~68 GB	Mac Pro or 4x RTX 4090	~10 tok/s

Installing via Ollama (Easiest Path)

ollama pull llama4:scout
ollama run llama4:scout

# Or with specific quantization
ollama pull llama4:scout-q4_K_M

llama.cpp Direct Setup

# Download GGUF
huggingface-cli download bartowski/Llama-4-Scout-17B-16E-Instruct-GGUF   --include "*Q4_K_M*"   --local-dir ./models/llama4-scout

# Run with full GPU offload (needs 55GB+ VRAM)
./build/bin/llama-server   -m ./models/llama4-scout/Llama-4-Scout-17B-16E-Instruct-Q4_K_M.gguf   --n-gpu-layers 999   --flash-attn   --ctx-size 32768   --host 0.0.0.0 --port 8080

Mac Studio M4 Max Setup (Single Machine)

The Mac Studio M4 Max with 128GB unified memory can run Llama 4 Scout at Q4_K_M with 16GB to spare for context. This is one of the few consumer machines that handles it without multi-GPU complexity.

# macOS with Metal acceleration
./build/bin/llama-server   -m ./models/llama4-scout-q4_k_m.gguf   --n-gpu-layers 999   --flash-attn   --ctx-size 65536   --threads 16

Performance Tips

• Use --flash-attn — reduces VRAM by 30% for long contexts on Scout
• Use --batch-size 512 for better throughput on multi-turn conversations
• Scout's MoE routing benefits from larger batch sizes — use --ubatch-size 512 alongside
• Temperature 0.7, top_p 0.9 is the sweet spot for general tasks

Llama 4 Scout vs Maverick: Which to Run?

Model	Total Params	Active Params	Experts	Best For
Scout	109B	17B	16	Home use, fast inference
Maverick	400B	52B	128	Higher quality, server use

For most home users, Scout hits the right balance of quality and speed. Check LLMHardware.io's Llama 4 benchmark results for measured token speeds on different GPUs.