Meta releases Llama 2 for commercial use

Meta released Llama 2, publishing model weights and code for pretrained and chat-tuned versions under a licence it described as free for both research and commercial use. This was the pivotal difference from the original Llama, whose weights had been released to approved researchers only in February 2023 and had then leaked onto the open internet within days, spreading well beyond Meta’s intended distribution regardless of its licence terms. Llama 2 made commercial use explicit and sanctioned rather than merely uncontrolled — with an exception, in the licence, requiring companies with more than 700 million monthly active users to seek a separate agreement from Meta.

Microsoft was named as Meta’s “preferred partner” for the release, with Llama 2 made available through Azure’s AI model catalog alongside Amazon Web Services and Hugging Face; Microsoft also produced versions optimised to run locally on Windows. The arrangement gave Meta broad distribution without operating its own inference service at the scale of OpenAI or Google, and gave Microsoft a credible open-weight alternative to hedge its dependence on OpenAI’s closed models.

Meta paired the release with safety material more substantial than the original Llama had carried: internal and external red-teaming, a published “transparency schematic” of known limitations, a responsible-use guide and an acceptable-use policy restricting certain applications. It also launched an Open Innovation AI Research Community for academic access and a “Llama Impact Challenge” for environmental and educational applications.

Llama 2 became the reference open-weight model of its period — the base other open developers fine-tuned, benchmarked against, and in several cases (including Mistral’s 7B model released two months later) explicitly tried to beat — and it anchored the argument that a well-resourced company willing to give weights away could keep pace with, or unsettle, labs selling access to closed models.