Created by pruning the MLP (feedforward) layers, reducing the size of Llama models while improving their performance.