Hybrid attention architecture, high-sparsity Mixture-of-Experts 80B model (active 3B). Currently supported for Mac only with MLX.