Facebook has announced a partnership with Microsoft to launch Llama 2, a next-generation open-source language model designed for commercial and research use. This major collaboration seeks to drive innovation and progress throughout the industry by making Llama 2 available for free to developers.
Facebook’s Commitment to Open-Source Infrastructure and AI Work
Facebook has a long history of open-sourcing its infrastructure and AI work, including PyTorch, ImageBind, Segment Anything, and Dino. This commitment to open-source technology has helped Facebook build better products, driving progress across the industry. By open-sourcing Llama 2, Facebook hopes to enable many more developers to build with new technology while improving safety and security.
Improved Architecture and Fine-Tuned Models
Llama 2 is equipped with pretrained and fine-tuned models with 7B, 13B, and 70B parameters. The model was pretrained on 40% more data than its predecessor, Llama 1, and boasts improvements to its architecture. For the fine-tuned models, Facebook collected more than 1 million human annotations and applied supervised fine-tuning and reinforcement learning with human feedback (RLHF) to achieve leading results on safety and quality.
Accessing Llama 2
Developers can download the models directly or access them via Microsoft’s Azure platform, along with Microsoft’s safety and content tools. An optimized version can also be run locally on Windows.
Mark Zuckerberg’s Statement
In a statement on his Facebook page, Mark Zuckerberg expressed his excitement about the possibilities that Llama 2 will unlock. He invited developers to share their ideas and projects, saying, “I’m looking forward to seeing what you all build!”