«Meta Llama» is a key product brand and research initiative of Meta Platforms, Inc.* (incorporated in 2004 in Delaware, USA, with headquarters in Menlo Park, California). This brand unites a family of large language models and associated infrastructure developed by the Meta AI* division. The name “Llama” has become synonymous with Meta’s* open artificial intelligence strategy, distinguishing this product line from the company's core social business.
The Llama series began with the release of LLaMA in 2023, which demonstrated the power of “intelligent scaling”—achieving high performance through data quality and efficient architecture rather than solely through massive parameter counts. This concept became the core development strategy for the company’s engineers. Among the most notable scientific breakthroughs was the introduction of Grouped-Query Attention (GQA), which optimizes attention computations for processing long sequences. However, the company’s true breakthrough in AI came with the release of Llama 3.1 in 2024; the series rapidly gained immense popularity, quickly fostering a vibrant community and driving infrastructure development. By 2025, Llama 4 achieved the next leap forward: Mixture-of-Experts (MoE), native multimodality (a unified architecture for text and images without external plugins), and an ultra-long context window of up to 10 million tokens.
Meta Llama has gained widespread recognition as a major driver of the open approach: the release of Llama 3.1 models and their high quality provided a powerful impetus for the transition from open models being used primarily by researchers, enthusiasts, and academia to their widespread adoption in real-world commercial products globally, proving the viability and competitiveness of open-source solutions at the enterprise level.
* The company Meta has been designated as an extremist organization in Russia.