On July 18, 2023, Meta released Llama 2, the successor to the LLaMA models from earlier that year. It came in 7-billion, 13-billion, and 70-billion-parameter sizes, with a chat-tuned variant, Llama 2-Chat, optimized for dialogue using reinforcement learning from human feedback. The accompanying paper reported that Llama 2-Chat outperformed other open chat models on most benchmarks tested.
The pivotal change was the license. The original LLaMA had been distributed for research use only, and its weights spread widely after a leak. Llama 2 was, in Meta’s words, free for research and commercial use, making it the first capable open model that companies could legally build products on without negotiating a separate license (subject to a clause restricting the very largest platforms by monthly active users). Meta launched it with Microsoft as a preferred partner, putting Llama 2 in the Azure AI model catalog and optimizing it to run locally on Windows, with availability also through AWS and Hugging Face.
Llama 2 turned the open-weights movement from a research phenomenon into a commercial one. A large ecosystem of fine-tuned derivatives, quantized builds, and products grew on top of it, and it set the template Meta would follow with Llama 3 and the frontier-scale Llama 3.1.