Meta has released Llama 2, a family of state-of-the-art open-access large language models, now fully integrated into Hugging Face. The release includes base models and fine-tuned chat versions with 7B, 13B, and 70B parameters, available under a permissive community license for commercial use.
The Llama 2 pretrained models boast significant improvements over Llama 1, including training on 40% more tokens, a longer context window of 4,000 tokens, and grouped-query attention for faster inference. The fine-tuned Llama 2-Chat models, optimized with Reinforcement Learning from Human Feedback (RLHF), perform competitively with ChatGPT on helpfulness and safety benchmarks.
Hugging Face provides a comprehensive ecosystem: models on the Hub, transformers integration, single-GPU fine-tuning examples, Text Generation Inference for production-ready serving, and Inference Endpoints. Users can request access via Meta's official form and log in with their Hugging Face account to start using the models.
The release marks a major step for open-source AI, offering a powerful alternative to closed-source chatbots. With permissive licensing and top-tier performance, Llama 2 is poised to drive widespread adoption in research and commercial applications.