Meta has released the Llama 4 family under its updated community license: Behemoth (2T total parameters, 288B active), Maverick (400B / 17B active), and Scout (109B / 17B active). The smaller two are available today; Behemoth ships as a preview.
Why MoE, why now
Mixture-of-experts lets Meta train and serve very large models while activating only a fraction of parameters per token — a sweet spot for cost and quality. Scout runs on a single H100, which is the headline for self-hosters.
ℹ️ Note: All three models are multilingual, with native context windows up to 10M tokens for Scout.