# Meta Llama 2 Launch

Llama 2 is a family of large language models released by Meta AI in July 2023, notable for making weights openly available for commercial use, which expanded the open-source AI ecosystem.

Llama 2 is a family of large language models (LLMs) developed by Meta AI, released on July 18, 2023, in partnership with Microsoft. It is the successor to the original Llama model and includes both foundation models and instruction-tuned chat versions. Unlike its predecessor, Llama 2 was made available with weights for broad commercial use, subject to an acceptable use policy, which significantly influenced the open-source AI ecosystem.

Llama 2 models are part of the broader [large-language-model](https://www.wikiprompt.org/wiki/large-language-model) landscape, building on advances in [transformer](https://www.wikiprompt.org/wiki/transformer) architectures and [deep-learning](https://www.wikiprompt.org/wiki/deep-learning) techniques. The release was seen as a major step toward democratizing access to powerful AI, providing a competitive alternative to proprietary models from companies like [openai](https://www.wikiprompt.org/wiki/openai) and [google-deepmind](https://www.wikiprompt.org/wiki/google-deepmind).

## Background

The rapid progress of large language models in the early 2020s, setting the stage for Llama, is intertwined with the success of [generative-ai](https://www.wikiprompt.org/wiki/generative-ai) tools. The release of [GPT-3](https://www.wikiprompt.org/wiki/gpt-3) demonstrated that scaling up model size could yield dramatic improvements in capabilities. In November 2022, OpenAI launched ChatGPT, which gained widespread attention and triggered a surge in AI research and investment.

Meta AI, the AI research arm of Meta, had been active in the field, with notable contributions such as the [transformer](https://www.wikiprompt.org/wiki/transformer) architecture developed by researchers at [google-deepmind](https://www.wikiprompt.org/wiki/google-deepmind) and others. Meta's chief AI scientist, Yann LeCun, publicly voiced skepticism about the hype around large language models, stating that they are best suited for assisting with writing tasks, reflecting Meta's focus on making AI more practical and accessible.

## Initial Release and Leak

The original Llama family ([Llama 1](https://www.wikiprompt.org/wiki/llama-1), stylized as LLaMA) was announced on February 24, 2023, through a blog post and a technical paper. It was available in sizes from 1 billion to 65 billion parameters, trained on publicly available data. Initially, model weights were only accessible to researchers on a case-by-case basis under a non-commercial license.

However, the weights were soon leaked online. In March 2023, a torrent of the weights was shared on 4chan and spread rapidly. Meta sent takedown requests to HuggingFace and GitHub for repositories hosting the model, but the leak enabled widespread private use and research. Many viewed the leak as a turning point for open AI, with comparisons to [Stable Diffusion](https://www.wikiprompt.org/wiki/stable-diffusion) accelerating innovation.

## Llama 2 Release

On July 18, 2023, in partnership with Microsoft, Meta announced Llama 2, the next generation of the Llama family. The company released three parameter sizes: 7 billion, 13 billion, and 70 billion parameters. Meta also released both foundation models and fine-tuned variants. Specifically, the chat fine-tunes were designed for conversational use.

The models were trained on 40% more data than the original Llama, but the architecture remained largely unchanged. Llama 2 models used [residual-network](https://www.wikiprompt.org/wiki/residual-network) blocks with attention mechanism, benefiting from [layer-normalization](https://www.wikiprompt.org/wiki/layer-normalization) and other innovations. The instruction fine-tuning used [RLAIF](https://www.wikiprompt.org/wiki/rlaif) - style techniques, combining supervised fine-tuning and human feedback.

## Commercial Availability and License

The most notable aspect of the Llama 2 release was the decision to provide weights for free use, for many commercial applications. Meta released the models under a custom license that allows use, reproduction, and modification for both commercial and research purposes, with some restrictions on use cases as defined by the policy.

Because of the acceptable use policy and other terms, many critics argued that Llama 2 was not true open source by the standards of the Open Source Initiative (OSI). Nevertheless, the availability of a powerful language model open for commercial use was a milestone, enabling smallers organizations and individuals to build upon the technology without relying on paid APIs.

Microsoft's partnership was notable, as it integrated the release with its [azure](https://www.wikiprompt.org/wiki/azure) infrastructure. The models were made available on the Azure AI model catalog, but they could also be downloaded directly, which increased accessibility.

## Model Architecture and Training

Llama 2 employed an architecture similar to its predecessor, but with adjustments in scaling. The model size ranged to 70 billion parameters, allowing it to capture more complex patterns. The training used a mix of sources, including Web pages, books, and other public data, with a focus on quality filtering.

A key aspect of the training process involves optimizing cross-entropy loss, using principles like [learning-rate-schedule](https://www.wikiprompt.org/wiki/learning-rate-schedule) and [model-pruning](https://www.wikiprompt.org/wiki/model-pruning). Meta also incorporated techniques such as [dropout](https://www.wikiprompt.org/wiki/dropout) and [weight-initialization](https://www.wikiprompt.org/wiki/weight-initialization) to improve performance.

The 7B, 13B, and 70B models are available, with the 70B model designed to run on multiple GPUs. Llama 2 was evaluated against contemporary models such as [GPT-3](https://www.wikiprompt.org/wiki/gpt-3) and [Chinchilla](https://www.wikiprompt.org/wiki/chinchilla), achieving competitive results on various benchmark tasks, including reasoning, and coding.

## Impact on Open-Source AI Ecosystem

Llama 2's release, with commercial access, significantly accelerated the growth of the open-source AI ecosystem. It provided an alternative to proprietary models, fostering a community of developers, researchers, and startups. Many companies, such as [AI21-Labs](https://www.wikiprompt.org/wiki/ai21-labs) and [inflection-ai](https://www.wikiprompt.org/wiki/inflection-ai), used Llama 2 as a foundation for their own models.

Furthermore, the release was accompanied by integration into cloud services like [amazon-web-services](https://www.wikiprompt.org/wiki/amazon-web-services), [google-cloud](https://www.wikiprompt.org/wiki/google-cloud), and [oracle-cloud](https://www.wikiprompt.org/wiki/oracle-cloud), making deployment accessible. Moreover, hardware vendors [amd](https://www.wikiprompt.org/wiki/amd), [intel](https://www.wikiprompt.org/wiki/intel), and [qualcomm](https://www.wikiprompt.org/wiki/qualcomm) optimized their chips to run Llama efficiently.

## Derivatives and Fine-tunes

The Llama 2 models were frequently fine-tuned for various applications. A notable example is Code Llama, a fine-tune with specialized code datasets. Released on August 24, 2023, initially in 7B, 13B, and 34B versions, and later a 70B version on January 29, 2024. Code Llama was trained on 500 billion tokens of code data, following on the foundation model. The [generative-ai](https://www.wikiprompt.org/wiki/generative-ai) community produced many other fine-tunes, such as models for creative writing, legal assistance, and medical advice.

The accessibility of model weights allowed researchers to apply techniques like [model-pruning](https://www.wikiprompt.org/wiki/model-pruning), [quantization](https://www.wikiprompt.org/wiki/quantization) and weights, making it possible for small AI companies andeducational institutions to study and adapt the technology.

## Response and Criticism

Reactions to Llama 2 were generally positive, but also concerning. Supporters praised its potential to democratize AI and foster innovation, noting that it could compete with commercial models like GPT-3.5. However, some critics pointed out that the licensed restrictions were still not entirely "open" as defined, thereby limiting their extent of freedom.

The acceptable use policy restricted applications in areas like surveillance, spamming, and malware. This was debated as an effort to prevent misuse, but also limited flexibility.

Some researchers also noted the environmental and resource concerns of training large models like Llama 2. Training required significant computing power, raising ethical and practical issues.

## Legacy and Future

Llama 2 set a new standard for open AI, leading to the later release of Llama 3 in April 2024. The release had a lasting effect on the industry, pushing major commercial labs to reconsider how models could be shared. It also encouraged the growth of the open AI collection as the basis for other products.

Llama 2's approach of giving weights with restrictions has been adopted by other developers to build AI assistants like Meta AI which is now built on Llama 3. As of late 2025, Llama is a family of models, and despite the emergence of new models, Llama 2 remains a foundational milestone.

## Summary

In sum, the Meta Llama 2 announcement was a significant event in the advancement of generative AI. By combining strong performance with permissive weights, what it said: Meta's decision helped shape the open AI ecosystem. The event is often cited as one of the edges that warmed the interest of commercial AI.

As of its release, Llama 2 marked a departure from the more restricted practices of other AI models, and its impact was widespread.

---
Source: https://www.wikiprompt.org/wiki/meta-llama-2-launch
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-12T16:24:48.428513+00:00