Why Llama 4 Scout (109B MoE) pressures system RAM
Llama 4 Scout (109B MoE) is Mixture-of-Experts: inference activates 17B active/token, but VRAM/RAM must usually hold the full ~109B expert set for fast routing. At Q4 the weight slab is ~61.3GB before KV (~0.04GB at 8K) and ~8GB OS/runtime overhead — totaling ~69.3GB raw, rounded to a 96GB kit. Stretching toward the full 10M-token window multiplies KV far faster than weights; that is the usual “I bought enough RAM for the model but still OOM” failure on Meta Llama MoE pages.
What RAM kit to buy
Buy a matched dual-channel DDR5 kit at 96GB for Llama 4 Scout (109B MoE) (EXPO/XMP only if stable). Avoid single-stick installs — local inference is bandwidth-sensitive when layers spill to host memory. Pair with 4x RTX 3090 / 4090 (96GB VRAM) or Apple Mac Studio (128GB Unified Memory) when staying in the Multi-GPU Workstation / Mac Studio tier, and keep 20–30% RAM free for the OS + browser.
Workload notes
Meta Llama-family models like Llama 4 Scout (109B MoE) have broad llama.cpp/Ollama support — prioritize stable JEDEC/EXPO kits over unproven XMP outliers for multi-hour serves. At 109B, Llama 4 Scout (109B MoE) sits in the large local-LLM band: Q4 on a strong GPU is realistic, FP16 usually is not on consumer cards. Release window noted as April 2025; always re-check the official source before buying hardware for a specific checkpoint.