FreeLLMMeta releases Llama 4 Scout and Maverick, 17B‑parameter MoE models with 10M context
Llama 4 Scout fits a single H100 GPU, while Maverick runs on a single H100 host, both using MoE routing for multimodal AI.
Meta announced two new mixture‑of‑experts (MoE) models: Llama 4 Scout and Llama 4 Maverick.
Both models have 17 billion active parameters. Scout uses 16 experts and fits on a single NVIDIA H100 GPU with Int4 quantisation.
Maverick employs 128 experts, fits on a single H100 host, and activates 17 B of its 400 B total parameters per token.
"best multimodal model in its class" is used to describe both Scout and Maverick in Meta’s release.
What it means for you
For developers, the models provide high‑quality multimodal performance on a single H100, lowering hardware costs for inference. Users of free or locally hosted models still rely on smaller, non‑MoE solutions, but the open‑weight release may inspire community‑driven equivalents.
Sources
- MoE Routing Strategies Explained: Top-K, Expert-Choice, Sigmoid Gating and Dynamic Routing (2026 Guide), MarkTechPost, 8 Nov 2026
- The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation, ai.meta.com (official), 5 Apr 2025
Written automatically from the sources above, then checked by our code: every number and name must appear in them, and no sentence may copy a source. Spotted a mistake? Report it.
Which AI can you use for free today?
Every free AI model, ranked by blind human votes, with the memory and speed it needs on your own GPU.