Llama 2 is trained on 40% more public data and can process twice as much context than Llama 1, according to Meta.
Big Tech firms Meta and Microsoft have teamed up to launch Llama 2, an open-source large language model from Meta that will feature on Microsoft’s Windows and cloud computing platform Azure.
The pair announced the collaboration on July 18, saying Llama 2 was made free for research and commercial use while also being optimized to run on Windows.
The announcement confirmed rumors from last week that said Llama 2 would be built for businesses and researchers to create applications on Meta’s AI tech stack.
Meta claimed Llama 2 was trained on 40% more publicly available online data sources and can process twice as much context compared to Llama 1.
The firm said Llama 2 outperforms many competitor open-source LLMs when it comes to coding, proficiency, reasoning, and performance on knowledge tests. However, Meta conceded it isn’t quite as efficient compared to its closed-source competitors, such as OpenAI’s GPT-4, according to one of its research papers
In a July 18 Instagram post, Meta CEO Mark Zuckerberg said Llama 2 “gives researchers and businesses access to build with our next generation large language model as the foundation of their work.”
Mark Zuckerberg with Microsoft CEO Satya Nadella. Source: Instagram
Meta said it was “blown away” by the demand for Llama 1 following the release of its limited version in February, which received over 100,000 requests for access. The model was soon leaked online by a user of the imageboard website 4chan.
Llama 1's figures, however, were far off from ChatGPT’s, which saw an estimated 100 million or more users sign up to use the model in the first three months, according to a February Reuters report.
With the partnership, Microsoft now backs two big players in the AI space, having invested a cumulative $13 million in OpenAI over the course of 2023, according to a January report by Fortune.
Meta’s decision to open source Llama was criticized by two United States senators in June, who claimed that the “seemingly minimal” protections in the first version of Llama potentially opened the doors for malicious users to engage in “criminal tasks.” (Source)