The new LLM features two models – one with 8 billion ... The development team used a standard decoder-only transformer architecture for Llama 3 but made key improvements over Llama 2.