Meta releases "Llama 3" — redefines open model performance, integrated into Meta AI assistant
On April 18, 2024, Meta released its next-generation model “Llama 3.” Billed as “the most capable openly available LLM to date,” it launched in 8B and 70B sizes. At the same time, Meta announced it had revamped its own assistant, “Meta AI,” to run on Llama 3 and integrated it into Facebook, Instagram, WhatsApp, and Messenger.
Details
- Two sizes released: Offered pretrained and instruction-tuned versions at 8B and 70B parameters, pretrained on a massive dataset of roughly 15 trillion tokens to strengthen reasoning capability
- Benchmarks: Achieved then-state-of-the-art performance across a wide range of industry benchmarks, pushing the capability of open models closer to GPT-4’s level
- Enhanced safety features: Introduced “Llama Guard 2” for improving input/output safety, “Code Shield” for code-related safety, and “CyberSec Eval 2” for cybersecurity evaluation
- Meta AI revamp: The “Meta AI” assistant — integrated across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta smart glasses — was rebuilt on Llama 3
- Distribution partners: Made available the same day across a wide range of platforms, including AWS, Databricks, Google Cloud, Hugging Face, Microsoft Azure, and NVIDIA NIM
What happened next
Llama 3 was expanded that July into the 405-billion-parameter-class “Llama 3.1,” as Meta further solidified its position as a provider of “open frontier models.” The integration of the Meta AI assistant also became a key driver behind the subsequent spread of the Llama series.