Meta unveils Llama 3, a cutting-edge language model elevating AI capabilities

Meta has unveiled the latest iteration of its open-source large language model Llama 3, introducing models trained with 8B and 70B parameters. These models are designed to support a wide range of use cases with improved reasoning, making them the best-in-class open source models in the industry. Integrated into the Meta AI assistant, these models aim to drive innovation in AI across various applications, development tools, and inference optimizations.

While the release of the models was initially anticipated in May, Meta has launched the first two models of the new Llama 3 generation earlier. These text-based models have been trained and fine-tuned with 8 billion and 70 billion parameters, showcasing advancements in reasoning, code generation, and instruction capabilities. They offer state-of-the-art performance across industry benchmarks and represent a significant improvement over the previous Llama 2 model.

In a comparative analysis with other models of similar parameter scales, such as Google’s Gemini and Anthropic’s Claude, Llama 3 has outperformed in the MMLU benchmark, demonstrating superior overall knowledge. Meta has conducted evaluations to showcase the model’s capabilities in various use cases, highlighting its effectiveness in tasks like advice-giving, brainstorming, topic classification, and coding skills, among others.

To train the Llama 3 model, Meta utilized a large dataset of more than 15T tokens from publicly available sources, seven times larger than the dataset used for Llama 2. The training data includes information in over 30 languages, covering a wider linguistic scope. Meta has also emphasized its commitment to responsible development of the model, introducing resources like Llama Guard 2, Code Shield, and CyberSecEval 2 to ensure safe usage.

The Llama 3 models have been integrated into the Meta AI assistant on various platforms like Facebook, Instagram, WhatsApp, and Messenger. They are also available for download and will soon be accessible through other cloud platforms and hardware providers. Meta has announced plans to introduce new capabilities, larger model sizes, and improved performance for Llama 3 in the coming months, demonstrating their ongoing commitment to AI innovation.