Meta's LLaMA-65B AI Model Weights Leak onto 4chan
The weight data for Meta's highly efficient LLaMA-65B language model leaks on 4chan, allowing unauthorized public downloads. High-speed downloaders quickly emerge online to distribute the AI model files.
The weight data for Meta's LLaMA-65B language model leaks on 4chan after an anonymous user posts torrent files and magnet links in an AI chatbot thread. LLaMA is a highly efficient large-scale language model that performs comparably to GPT-3 but requires only a single GPU to operate, making it highly desirable for consumer-level hardware.
Following the initial 4chan leak, developers quickly create high-speed downloaders on GitHub that allow users to grab the 65 billion parameter weights at speeds up to 40MB/s. Additionally, unauthorized contributors attempt to add the leaked magnet links directly to Meta's official LLaMA GitHub repository to facilitate broader distribution of the model data.
This unauthorized release undermines Meta's controlled distribution strategy, which previously required researchers to contact the company directly to access the neural network weights. The leaked data includes not only the massive 65B model but also the smaller 7B, 13B, and 30B variants, sparking widespread discussions across online platforms about the implications of open-access AI.