Llama 2 is a family of large language models (LLMs) developed and released by Meta AI in July 2023. It is the successor to the original Llama model and is designed to be more accessible and commercially usable than its predecessor. Llama 2 includes models with 7, 13, and 70 billion parameters, each available in both a foundation version and a version fine-tuned for chat applications. The models are released with weights, allowing for broad use, but under a license that restricts certain applications, leading to debate about whether they qualify as open source.
The architecture of Llama 2 largely follows that of the original Llama, but the training data was increased by 40%. This expansion aimed to improve performance and generalization across various natural language processing tasks. The release of Llama 2 marked a significant step in Meta's strategy to compete with other LLM providers, such as OpenAI and Google DeepMind, by offering a powerful yet accessible alternative.
Background
The development of Llama 2 occurred against the backdrop of rapid advancements in large language models, particularly after the release of GPT-3 and the subsequent popularity of ChatGPT. These models demonstrated that scaling up neural networks could lead to significant improvements in capabilities, sparking intense interest in the field. Meta's Chief AI scientist, Yann LeCun, noted that LLMs are particularly effective for writing assistance, positioning Llama as a tool for creative and professional writing tasks.
The original Llama, released in February 2023, was initially available only to researchers under a non-commercial license. However, its weights were leaked online in March 2023, leading to widespread unofficial distribution. This leak, while controversial, also highlighted the demand for accessible LLMs and influenced Meta's decision to release Llama 2 with a more permissive license.
Versions
Initial Release
The first version of Llama, sometimes referred to as Llama 1, was announced on February 24, 2023. It was trained on publicly available data and came in sizes ranging from 1 billion to 65 billion parameters. The 13B parameter model outperformed the much larger GPT-3 on many benchmarks, while the 65B model was competitive with state-of-the-art models like PaLM and Chinchilla. Access to the weights was restricted to academic researchers and industry labs, but the inference code was released under the open-source GPLv3 license.
#### Leak
On March 3, 2023, a torrent containing Llama's weights was uploaded to 4chan, and links quickly spread through online AI communities. Meta filed takedown requests with HuggingFace and GitHub, but the leak had already facilitated widespread use. Commentators compared the situation to the release of Stable Diffusion, suggesting that open access could spur innovation and tool development, despite concerns about misuse.
Llama 2
Llama 2 was announced on July 18, 2023, in partnership with Microsoft. It was released in three sizes: 7B, 13B, and 70B parameters. The architecture remained similar to Llama 1, but the training data was increased by 40%. Unlike its predecessor, Llama 2's weights were made available for commercial use, subject to an acceptable use policy. This policy prohibits certain applications, leading the Open Source Initiative to dispute Meta's characterization of Llama as open source.
Llama 2 also served as the basis for Code Llama, a specialized version fine-tuned for code generation. Code Llama was released in August 2023 and January 2024, with versions ranging from 7B to 70B parameters, trained on additional code-specific datasets.
Llama 3
On April 18, 2024, Meta released Llama 3 with 8B and 70B parameter sizes. These models were pre-trained on approximately 15 trillion tokens of publicly available text, with instruction fine-tuning on over 10 million human-annotated examples. Llama 3 70B outperformed Gemini Pro 1.5 and Claude 3 Sonnet on many benchmarks. The models demonstrated that performance continues to improve even when trained on data far beyond the Chinchilla-optimal amount, with the 8B model showing log-linear scaling up to 15 trillion tokens.
Meta also announced plans to make Llama 3 multilingual and multimodal, and to increase its context window. The latest version, Llama 4, was released in April 2025, and in April 2026, Meta Superintelligence Labs introduced Muse Spark as a replacement for Llama.
Impact and Reception
The release of Llama 2 and subsequent versions has had a significant impact on the AI landscape. By providing open weights, Meta enabled researchers and developers to fine-tune models for specific applications, fostering a vibrant ecosystem of tools and derivatives. However, the licensing terms have sparked debates about the definition of open source, with critics arguing that the acceptable use policy restricts freedom.
Llama models have been widely adopted in both academia and industry, with applications ranging from chatbots to code generation. The models' performance, particularly in relation to their size, has made them attractive for organizations with limited computational resources. As of 2025, Llama remains one of the most prominent open-weight LLM families, influencing the development of subsequent models and shaping discussions about AI accessibility and governance.