one year on
Meta releases Llama 2, an open-source large language model free for commercial use
The next generation of Meta's large language model is available for free research and commercial use, with Microsoft as preferred partner, challenging the closed-source approach of rivals like OpenAI.
Meta today released Llama 2, the next generation of its large language model, making it free for research and commercial use. The model comes in 7 billion, 13 billion, and 70 billion parameter sizes, with a chat-tuned variant called Llama 2-Chat. It was trained on two trillion tokens, nearly double the 1.4 trillion tokens used for Llama 1.
Meta is partnering with Microsoft, which becomes the preferred partner for Llama 2. The model is available on Azure, AWS, and Hugging Face. Microsoft CEO Satya Nadella announced the expanded partnership at Microsoft Inspire. Qualcomm says it’s working to bring Llama 2 to Snapdragon devices in 2024.
While Meta emphasizes the safety and responsibility of the release—including red-teaming, a responsible use guide, and an acceptable use policy—Meta says it is releasing Llama 2 under a license and acceptable-use policy that prohibit certain use cases. The company claims human evaluators found Llama 2 roughly as helpful as ChatGPT across a set of roughly 4,000 prompts designed to probe for helpfulness and safety.
The release marks a shift from Llama 1, which was gated by request and later leaked online. Now Meta fully embraces open source, hoping to spur innovation and improve safety through community scrutiny. Meta says the model still carries biases along some axes and can be overly cautious on some prompts. The debate around what counts as truly ‘open source’ for AI models is reignited with this release.
The record
One year later — open only if you can handle spoilers
Llama 2 became one of the most widely used open-weight base models, spurring a wave of fine-tuned variants and commercial products. It cemented the 'open vs. closed' debate as a defining axis of the AI industry, with Meta positioning itself as the open champion against OpenAI and Google. The 700 million monthly active user license carve-out was later relaxed, but the commercial use license still faced criticism from open-source purists.