# Llama 2 Open-Source Release

Meta released Llama 2 on July 18, 2023, as an open-weights large language model for research and commercial use, marking a shift from the restricted Llama 1.

Llama 2 is a family of large language models ([LLMs](https://www.wikiprompt.org/wiki/large-language-model)) released by [Meta AI](https://www.wikiprompt.org/wiki/meta-ai) on July 18, 2023, in partnership with [Microsoft](https://www.wikiprompt.org/wiki/microsoft). Unlike the first version of Llama, which was restricted to academic researchers under a non-commercial license, Llama 2 was made available to the public with model weights and a license that permitted many commercial use cases. The release included both foundation models and instruction-tuned chat versions, in sizes of 7, 13, and 70 billion parameters.

The launch of Llama 2 was a significant event in the field of [generative AI](https://www.wikiprompt.org/wiki/generative-ai), as it provided a competitive open-weights alternative to proprietary models like [GPT-3](https://www.wikiprompt.org/wiki/gpt-3) and [ChatGPT](https://www.wikiprompt.org/wiki/chatgpt). It also sparked debate about the meaning of "open source" in AI, since Meta's license included usage restrictions that some argued deviated from the Open Source Definition. The release influenced subsequent developments in open-weights modeling, including Meta's own Llama 3 and other models from companies like [Mistral AI](https://www.wikiprompt.org/wiki/mistral-7b-release) and [AI21 Labs](https://www.wikiprompt.org/wiki/ai21-labs).

## Background

The development of Llama 2 came against a backdrop of rapid advances in large language models. After the release of [GPT-3](https://www.wikiprompt.org/wiki/gpt-3-2020) in 2020, researchers found that scaling models to larger sizes often led to dramatic improvements in capabilities, such as reasoning and code generation. The launch of ChatGPT in November 2022 brought LLMs to a broad public audience and triggered an industry-wide race to develop and improve these models.

Meta's initial response was cautious. In an interview, Chief AI scientist [Yann LeCun](https://www.wikiprompt.org/wiki/yann-lecun) noted that LLMs were best suited for assisting with writing, rather than serving as general-purpose reasoning engines. However, the success of ChatGPT and the subsequent [leak of the first Llama model](https://www.wikiprompt.org/wiki/llama-leak-march-2023) in March 2023 led Meta to reassess its strategy. By releasing Llama 2 openly, Meta aimed to compete while also fostering an ecosystem of external developers and researchers.

## Release announcement

Meta announced Llama 2 on July 18, 2023, in a blog post and a paper titled "Llama 2: Open Foundation and Fine-Tuned Chat Models." The release was accompanied by a partnership with Microsoft, which made Llama 2 available on its [Azure](https://www.wikiprompt.org/wiki/azure) cloud platform and [Microsoft](https://www.wikiprompt.org/wiki/microsoft)'s Windows operating system. Meta also provided a responsible use guide and a community license that allowed commercial use, with stipulations such as not using the model to improve other large language models.

The announcement highlighted that Llama 2 was trained on 40% more data than its predecessor, Llama 1, and that the chat versions were fine-tuned using reinforcement learning from human feedback (RLHF), a technique similar to that used for ChatGPT. Meta reported that the 70B chat model outperformed ChatGPT on some benchmarks, though these claims were later debated.

## Model architecture and training

Llama 2 retains the same transformer-based architecture as Llama 1, but with adjustments in training. The models use the same tokenizer and parameter counts, but were trained on 2 trillion tokens of publicly available web data, compared to 1.4 trillion for Llama 1. The training data included mixtures of English-language texts, code, and other sources.

The three sizes - 7B, 13B, and 70B - were chosen to accommodate different hardware and use cases. The 7B and 13B models can run on consumer GPUs, while the 70B model requires more substantial computational resources. Meta also released debug and quantized versions to facilitate local inference.

For the chat fine-tunes, Meta used a multi-stage process: supervised fine-tuning on human demonstrations, then RLHF with proximal policy optimization, followed by rejection sampling. The company also implemented safety measures in the chat models to reject harmful prompts, though these were not always effective.

## Open license and controversy

The license for Llama 2 was hailed as a step forward from Llama 1, which had been limited to academic researchers. However, the license was not approved by the Open Source Initiative (OSI), which maintains the official definition of open source. The OSI argued that the restrictions on using Llama 2 to improve other LLMs and the user limits for large companies violated the principles of open source, which require no field-of-use restrictions and no discrimination against persons or groups.

Meta continued to use the term "open source" to describe Llama 2, leading to criticism from some developers and researchers who felt the word was being diluted. This debate intensified with later releases, as Meta's Llama 3 and other models used similar licenses.

Despite the controversy, the availability of Llama 2's weights allowed a wide range of applications. Researchers could fine-tune the model for specific domains, and startups could build commercial products without paying license fees, so long as they complied with the acceptable use policy.

## Reception and impact

Initial reactions to Llama 2 were mixed. Many praised Meta for releasing a powerful model with fewer restrictions, comparing it to the impact of [Stable Diffusion](https://www.wikiprompt.org/wiki/stable-diffusion-public-release) in image generation. Developers quickly created fine-tunes for coding, chat, and other purposes. The model was integrated into platforms like [hugging-face](https://www.wikiprompt.org/wiki/hugging-face) (though not a provided slug) and [AWS](https://www.wikiprompt.org/wiki/amazon-web-services) and [Google Cloud](https://www.wikiprompt.org/wiki/google-cloud).

However, some critics noted that Llama 2's performance, especially for the 70B model, did not always match that of proprietary models like [GPT-4](https://www.wikiprompt.org/wiki/gpt-4-release) or [PaLM 2](https://www.wikiprompt.org/wiki/palm). Independent benchmarks showed that Llama 2 70B was competitive with [GPT-3.5](https://www.wikiprompt.org/wiki/gpt-3) in many tasks but lagged behind the state of the art in reasoning and factuality.

The release also had an impact on the broader AI community. It prompted other companies, such as [Mistral AI](https://www.wikiprompt.org/wiki/mistral-7b-release), to release open-weights models, and it contributed to the growth of a decentralized ecosystem of LLM deployment. In addition, Llama 2's availability on [Microsoft](https://www.wikiprompt.org/wiki/microsoft) Azure made it easier for enterprises to adopt open-weights AI.

## Comparison with Llama 1 and successors

Llama 2 differed from Llama 1 in several key ways: it was open for commercial use, it included chat fine-tunes, and it was trained on more data. However, the architecture was largely unchanged, and some experts noted that the improvements were incremental rather than revolutionary.

Subsequent releases built on Llama 2. In August 2023, Meta released Code Llama, a fine-tune of Llama 2 specialized for code generation, in 7B, 13B, 34B, and 70B sizes. In April 2024, Meta introduced [Llama 3](https://www.wikiprompt.org/wiki/meta-llama-3-2024), with 8B and 70B sizes, which used a larger training dataset and showed improved performance. Llama 3 was followed by [Llama 4](https://www.wikiprompt.org/wiki/meta-llama-4-launch) in April 2025, which became part of Meta's AI assistant. By 2026, Meta Superintelligence Labs released Muse Spark as a successor.

## Legacy

Llama 2's legacy lies in its role as a catalyst for the open-weights movement in AI. It demonstrated that powerful models could be shared widely while still maintaining safety controls, and it provided a practical alternative to closed, API-only models. The model also influenced discussions about the definition of open source in AI, a debate that continues as the field evolves.

For many developers and researchers, Llama 2 became a baseline for experimentation and innovation. Its release lowered the barrier to entry for working with large language models, enabling a wave of fine-tunes and applications that might not have existed otherwise. Even as newer models emerged, Llama 2 remained a reference point for understanding the trade-offs between openness, capability, and safety in AI development.

## External links

- [Wikipedia: Llama 2](https://en.wikipedia.org/wiki/Llama_2)

---
Source: https://www.wikiprompt.org/wiki/llama-2-open-source-release
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-11-06T12:15:51.840307+00:00