Meta's mixture-of-experts open-weight family — Scout and Maverick, both with 17B active parameters. Available variants: Scout 109B MoE (active 17B)Maverick 400B MoE (active 17B).
Llama 4 is Meta's first mixture-of-experts family: Scout (109B total, 16 experts) and Maverick (400B total, 128 experts) both route each token through about 17B active parameters, so they generate at roughly the speed of a 17B dense model but need memory for the whole expert pool. Scout at Q4 is about 60 GB, Maverick about 220–245 GB. There is no 8B or 70B Llama 4 — those sizes are Llama 3.1 / 3.3, which remain the better fit for a single 8–24 GB card.
Hardware tiers for running Llama 4
🧪
Minimum (Scout at Q4, GPU + expert offload)
Scout's ~60 GB of Q4 weights do not fit a 24 GB card, but only 17B parameters are active per token, so keeping attention and shared layers on the GPU and the routed experts in system RAM stays usable for chat. A ~2-bit quant (~40 GB) cuts the RAM need at a quality cost.
VRAM24 GB
System RAM64–96 GB DDR5
ThroughputRoughly 10–20 tok/s with llama.cpp expert offload — scales with system-RAM bandwidth
*Price sourced from Amazon.com. Price and availability subject to change.
🏆
Recommended (Scout at Q4, GPU + lighter expert offload)
Scout's ~60 GB of Q4 weights do not fit a 32 GB card either, so some routed experts still live in system RAM. The extra 8 GB over a 24 GB card keeps more of them resident and leaves room for longer context. Fully in VRAM, Scout needs about 64 GB at Q4 with a short context: two 32 GB RTX 5090s, or one 96 GB RTX PRO 6000 Blackwell with room for long context.
VRAM32 GB
System RAM64–128 GB DDR5
ThroughputFaster than a 24 GB card because fewer experts sit in system RAM; still bounded by RAM bandwidth
*Price sourced from Amazon.com. Price and availability subject to change.
⚡
Luxury (Maverick 400B at Q4)
Maverick is beyond any consumer GPU. It fits a 256 GB+ Mac Studio M3 Ultra, three or more 96 GB RTX PRO 6000 cards, or a server with a 24–32 GB GPU plus 256 GB+ of RAM for expert offload.
VRAM96 GB
System RAM256 GB+ DDR5, or 256–512 GB unified memory
ThroughputNeeds ~220–245 GB for Q4 weights before any context
Llama 4 switched to mixture-of-experts. Llama 3.1 / 3.3 are dense 8B, 70B and 405B models; Llama 4 ships as Scout (109B total, 17B active) and Maverick (400B total, 17B active), with native image input and much longer context (up to 10M tokens on Scout). Each token touches only 17B parameters, but the whole model still has to fit in memory.
Can I run Llama 4 Scout on a single 24 GB GPU?
Yes, with expert offload. Scout's Q4 weights are about 60 GB, so a 24 GB card keeps attention and the shared layers while the routed experts sit in 64 GB+ of system RAM. Aggressive ~2-bit quants shrink the file to roughly 40 GB at a quality cost. Fully in VRAM it needs about 64 GB — two RTX 5090s or one 96 GB workstation card.
Does Llama 4 Maverick run locally?
Only on very large memory pools. At Q4 Maverick needs roughly 220–245 GB, which means a 256 GB+ Mac Studio M3 Ultra, several 96 GB workstation cards, or a server pairing a 24–32 GB GPU with 256 GB+ of system RAM for expert offload.
Which Llama should I run on a 12–24 GB card?
Llama 3.1 8B fits any 8–12 GB card at Q4. Llama 3.3 70B needs about 40 GB at Q4, so two 24 GB cards or one 48 GB card. Llama 4 Scout on a 24 GB card works only with its experts offloaded to system RAM.
More guides & deep dives from the SpecPicks archive
