The new generation large model Llama 3, open-sourced by Meta, has released only the 8B and 70B versions so far. However, it has already outperformed large models such as Claude 3 Sonnet, Mistral Medium, and GPT-3.5 in evaluation results. Compared to Llama 2, it has made improvements in parameter scale, training datasets, model architecture (GQA), performance, multilingual support, inference, and code generation.