Llama 4 Scout and Maverick, Open MoE and Long Context

By Satwik · June 18, 2026

Meta released Llama 4 in April 2025 as a family of natively multimodal, mixture-of-experts models, led by Scout and Maverick. Both used MoE architectures that activate only a fraction of total parameters per token, aiming for strong quality at lower inference cost, and both were trained to handle image and text together. The eye-catching claim was Scout's very long context window, marketed at up to 10 million tokens, a dramatic jump that, whatever its practical usable length, signaled where Meta wanted the open ecosystem to go.

The release mattered as the open-weights counterweight to the closed frontier. Llama had become the default foundation for a huge amount of downstream fine-tuning and agent building, so Llama 4's architecture choices, MoE and long context, effectively set defaults for the open community. Reception was mixed: some benchmark and evaluation controversies tempered the launch, and the very long context claim drew scrutiny about real-world effectiveness versus headline numbers, a reason to stay measured about the specifics.

For agentic security, open weights carry the familiar structural point: guardrails ship as removable defaults, and a widely deployed base model becomes the substrate for countless agents whose safety depends on integrators, not Meta. The long context window enlarges the indirect-injection surface, since more retrieved or pasted content means more places to hide adversarial instructions, and MoE routing adds a subtler research question about whether expert selection can be probed or manipulated. Llama 4's lasting role is as infrastructure for the open agent stack, which makes its architecture and its known limitations the shared inheritance, and shared risk, of everyone building autonomous systems on top of it.