Hybrid SSM-Attention language model on Apple Silicon with MLX — interleaving Mamba-2 and Transformer for efficient inference
