The Llama 4 Herd
Meta's Llama 4 Scout and Maverick introduced open-weight natively multimodal mixture-of-experts models, distilled from the larger Behemoth teacher model.
How far can open-weight multimodal mixture-of-experts models support inspectable agent research before capability gaps with closed frontier models dominate?
Llama 4 is older than the July 2026 frontier drops, but it belongs in the ASI map because it anchors the open-weight side of the capability race. Meta introduced Scout and Maverick as natively multimodal mixture-of-experts models.
Technical signal
Llama 4 Maverick uses 17B active parameters and 400B total parameters, with alternating dense and MoE layers. Meta also described Behemoth as a teacher model for the released systems.
ASI relevance
Open-weight frontier-adjacent models matter because they let outside labs inspect, adapt, distill, and build agents without waiting for closed APIs. That widens the experimentation surface for recursive and multi-agent systems.
Source trail: Meta announced Scout and Maverick on X.