Meta Llama 3 open-access LLM released on Hugging Face with 8B/70B, includes Llama Guard 2 — attribution required

AI Impact Summary

Meta has released Llama 3 as open-access models in two sizes (8B and 70B) with base and instruct-tuned variants, plus Llama Guard 2 for safety; all are available on Hugging Face. The 8B model uses Grouped-Query Attention and an expanded 128K-token vocabulary, enabling longer context and improved multilingual capability, while training reportedly used over 15 trillion tokens. Llama 3 licenses require explicit attribution and naming derivatives starting with 'Llama 3' and including 'Built with Meta Llama 3', affecting how teams package and distribute models or derivatives. The release with Hugging Face Transformers integration lowers barriers for rapid experimentation and production pilots, but 70B incurs substantial memory needs and deployment considerations, while 8B remains accessible on consumer GPUs.