Llama 3 expands Meta's openly available model family

Meta released the first Llama 3 models today in versions with eight billion and seventy billion parameters. Developers can download the instruction-tuned and pretrained models under Meta's license or access them through several cloud and hardware partners.

The company says training used more than fifteen trillion tokens from publicly available sources, with additional filtering and safety work. Meta is also placing an updated Meta AI assistant in Facebook, Instagram, WhatsApp and Messenger, supported by live search results.

Downloadable weights let researchers and companies run, inspect and adapt a capable model on infrastructure they control. That accessibility can support specialized tools and wider experimentation, although the license carries conditions and open availability does not reveal the training data. Larger multimodal and multilingual models are still in development. For now, Llama 3 increases pressure on closed-model providers by making customization and local deployment central parts of the competitive argument.