Llama 3 was released April 18, 2024 in 8B and 70B sizes. Trained on 15T tokens (7x Llama 2). The 70B model scored 82% on MMLU and 81.7% on HumanEval, approaching GPT-4 performance. Used a Grouped Query Attention (GQA) architecture and a 128K token vocabulary.
Modality
text
Capabilities1
tool-use