Llama 3: Definition and Examples

Llama 3 (Large Language Model Meta AI, 3rd generation) is a family of language models published by Meta in April 2024. It represents a significant leap over Llama 2, with performance rivaling proprietary models like GPT-4 and Claude on many benchmarks. Llama 3 is available in several sizes — notably 8B and 70B parameters — allowing it to adapt to various use cases, from deployment on a personal computer to the most demanding cloud infrastructures.

Meta's philosophy with Llama 3 is based on open source: the model weights are freely downloadable, allowing anyone to use, fine-tune, or integrate them into commercial applications (under a permissive license). This approach has catalyzed an entire ecosystem of tools, adaptations, and derivative models created by the community, making Llama 3 one of the most widely adopted open-source models in the world.

On the technical side, Llama 3 builds on an optimized Transformer architecture, an improved tokenizer (128K token vocabulary), and training on a massive corpus of over 15 trillion tokens. The model excels in reasoning, code generation, instruction following, and multilingual understanding. Meta has also released Llama 3.1 (with a 405B parameter version) and Llama 3.2 (integrating multimodal vision capabilities and lightweight versions for edge computing), solidifying Llama 3 as an ever-evolving platform.

For prompt engineering practitioners, Llama 3 offers the major advantage of being runnable locally or on private infrastructure, ensuring full data control. Its structured prompt format (with system, user, and assistant role tags) is compatible with advanced prompting techniques such as few-shot, chain-of-thought, and RAG.