About
Llama 3 is Meta's third generation of large language models, introduced on April 18, 2024. The family includes 8B and 70B parameter models in base (pretrained) and instruction-tuned (Instruct) variants. The models use a decoder-only transformer architecture with grouped-query attention (GQA), a 128K-token tokenizer vocabulary, and were pretrained on over 15T tokens. Weights are released under the Meta Llama 3 Community License.