Meta has unveiled its latest open-source AI model, Llama 3.1, designed to significantly benefit Indian startups by offering advanced synthetic data generation capabilities. This new model addresses the challenge of limited high-quality real-world data in local languages.
Ragavan Srinivasan, Meta’s Vice President of Product Management, emphasized that Llama 3.1 serves as a versatile base model that companies and developers can customize to better capture language and cultural nuances, including languages such as Hindi, Kannada, and Marathi.
Llama 3.1, featuring an impressive 405 billion parameters, aims to bridge the performance gap with leading closed-source models like OpenAI’s GPT-4 and Anthropic’s Claude 3.5 Sonnet. It was trained on 15 trillion tokens using 16,000 Nvidia H100 GPUs and supports synthetic data generation for fine-tuning and understanding linguistic subtleties.
Meta has also revised its commercial license to allow developers to generate synthetic data from Llama 3.1 for training other proprietary models. This model distillation capability enables large models to impart their intelligence to smaller models, reducing the need for extensive real-world data training.
Experts view Meta’s strategic move as a way to optimize price and performance for specific applications based on Llama’s broad foundation. Hemant Mohapatra, a partner at Lightspeed, commented, “Meta made a strategic bet. You can now achieve the specificity of small language models (SLMs) for niche use cases while leveraging Llama’s extensive knowledge base.”
Despite the high cost of deploying the 405B model for large-scale enterprise applications, Meta expects many organizations to prefer smaller versions, such as the 70B and 8B models, which will be available on the cloud. Srinivasan anticipates improvements in efficiency and optimization as part of the typical evolution of open-source software.
Llama’s distinctive model distillation feature is a notable advantage, not commonly found among proprietary closed vendors. With 300 million downloads globally, Llama’s various versions have made a significant impact, with India ranking among the top markets for Meta’s models.
While Meta’s AI assistant is gaining traction and competing with ChatGPT in popularity, experts note that surpassing ChatGPT remains challenging. Arun Chandrasekaran, a vice president and analyst at Gartner, pointed out, “ChatGPT continues to lead as the most popular conversational AI chatbot due to its higher accuracy, multi-modality, and extensibility.” He also noted that while Meta AI’s usage might expand within apps like Instagram, WhatsApp, and Facebook, OpenAI retains a strong presence with its versatile, general-purpose chatbot.
