The Llama 4 series is the first to use a “mixture of experts (MoE) architecture,” where only a few parts of the neural ...