As our first quantized models in this Llama category, these instruction-tuned models retain the quality and safety of the original 1B and 3B models, while achieving 2-4x speedup.