Zebra-Llama: Towards Efficient Hybrid Models