Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Models

RNN models Hawk and Griffin show superior performance and efficiency compared to Transformers in language modeling tasks.