Meta has released Llama 3.1 405B, its latest large language model (LLM). According to Facebook’s parent company, the new model is the world’s largest open-source LLM. It builds on its predecessor with new applications and abilities, which include synthetic data generation and the ability to train smaller models. In addition to introducing its 405B model, Meta has released upgraded versions of its Llama 8B and Llama 70B models.
Meta’s founder and CEO Mark Zuckerberg said that Llama 3.1’s predecessor, Llama 2, was only comparable to an older generation of LLMs. Its latest model, he argued, leads on several of its rivals. “Starting next year, we expect future Llama models to become the most advanced in the industry,” said Zuckerberg. “But even before that, Llama is already leading on openness, modifiability, and cost efficiency.”
Llama 3.1 builds on previous generation of LLMs
Meta said that Llama 3.1 will expand context length to 128K and add support across eight languages. The latest model will also support advanced use cases, including long-form text summarisation, multilingual conversational agents, and coding assistants. Additionally, Meta optimised its full training stack and pushed the model training to over 16,000 H100 GPUs to enable training runs at a massive scale.
The social technology company aims to advance Llama 3.1’s capabilities by delivering additional components, including a reference system to support development. It also plans to empower developers with tools to build custom agents and new agentic behaviours.
In line with responsible AI development, Meta is enhancing its offerings with new security and safety tools, such as the Llama Guard 3 multilingual safety model and Prompt Guard prompt injection filter. Additionally, Meta has issued a request for comments on the Llama Stack API, a standardised interface designed to simplify the integration of Llama models into third-party projects.
More to come for LLM’s corporate ecosystem
The social media giant said that its ecosystem of partner platforms for Llama 3.1 is ready for launch, with more than 25 partners already on board. Corporate partners include Cloudflare, Amazon Web Services (AWS), NVIDIA, Databricks, Groq, Dell, Azure, Google Cloud, and Snowflake.
“The Llama 3.1 model is particularly exciting, as it is released in a higher precision (bfloat16), incorporates function calling, and adds support across 8 languages,” said Cloudflare in a statement shortly after Llama 3.1’s announcement. “Having multilingual support built-in means that you can use Llama 3.1 to write prompts and receive responses directly in languages like English, French, German, Hindi, Italian, Portuguese, Spanish, and Thai.”
Read more: Meta AI cancels European launch
