Meta has unveiled the Llama 3.1 collection of models, including the 405B parameter version.
The 405B parameter version, is being touted as the first frontier-level open-source AI model. This release comes with a strong emphasis on responsible AI development and expanded safety measures.
The Llama 3.1 release features extended context length to 128K tokens and support for eight languages. Meta has also introduced new safety tools, including Llama Guard 3 for multilingual input and output moderation, Prompt Guard to protect against prompt injections, and CyberSecEval 3 for cybersecurity risk assessment.
Meta's approach to responsible AI development encompasses pre-deployment risk assessments, safety evaluations, and extensive red teaming with internal and external experts. The company has collaborated with partners like AWS, NVIDIA, and Databricks for safe distribution and has openly shared model weights, recipes, and safety tools.
Thorough evaluations have been conducted in critical areas such as cybersecurity, chemical and biological weapons risks, and child safety. Meta reports no significant increase in malicious actor capabilities using Llama 3.1 405B in these areas.
Mark Zuckerberg, CEO of Meta, emphasised the importance of open-source AI, stating it "helps ensure that more people around the world can access the opportunities that AI provides, guards against concentrating power in the hands of a small few, and deploys technology more equitably."
Meta's release of Llama 3.1, particularly the 405B model, represents a step forward in open-source AI development. By combining advanced capabilities with comprehensive safety measures and transparency, Meta aims to set a new standard for responsible AI innovation in the industry.