Llama 2

Llama 2 is a family of large language models released by Meta in July 2023 as the successor to the original LLaMA model. It was developed for general-purpose language generation, dialogue, coding assistance, summarization, question answering, and other text-based artificial-intelligence applications. The release included pretrained models and instruction-tuned chat variants, allowing developers to use the models either as foundations for further customization or as conversational systems with comparatively limited additional engineering. The model family was offered in 7-billion, 13-billion, and 70-billion-parameter versions. The larger models were intended to provide stronger reasoning and language performance, while the smaller versions were designed for more accessible deployment on constrained infrastructure. Llama 2 used a 4,096-token context window and was trained on a larger publicly described data mixture than the first LLaMA release. Meta also published model documentation, responsible-use guidance, evaluation information, and a license that permitted many commercial applications subject to stated conditions. Llama 2 became a major part of the open-model ecosystem because Meta made model weights available for download rather than limiting access to a hosted application or an application programming interface. Microsoft was announced as a preferred cloud partner at launch, and the models were distributed through channels including Microsoft Azure and other technology platforms. This broadened access for developers and enterprises that wanted to run, fine-tune, or integrate language models using their own infrastructure or a third-party cloud service. The Llama 2 brand covered several related offerings rather than a single application. The base pretrained models were intended for developers who wanted to perform their own adaptation. Llama 2-Chat models were fine-tuned for dialogue and aligned using supervised fine-tuning and reinforcement learning from human feedback. Meta described the chat models as suitable for assistant-like interactions but cautioned that they could still produce inaccurate, biased, unsafe, or otherwise undesirable output. Llama 2 was positioned between proprietary commercial systems and fully unrestricted open-source software. Meta called the release open and made the weights broadly available, but its license contained conditions and did not amount to an unrestricted open-source license under every commonly used definition. The license included attribution requirements and additional provisions for very large services, while Meta separately published an acceptable-use policy. These terms became part of the product's competitive identity and generated debate among developers and open-source advocates. The release helped establish Meta as a prominent supplier of foundation models and encouraged a broad ecosystem of hosted inference providers, fine-tuning tools, local deployment frameworks, and derivative models. Although later generations of the Llama family superseded Llama 2 in Meta's product roadmap, Llama 2 remains historically important as one of the releases that accelerated adoption of downloadable large language models for commercial and research development.

Llama 2 originated within Meta's Llama research program, which had produced the original LLaMA family as a set of large language models intended to support research. The first release attracted substantial attention because its model weights were made available to approved researchers, but access and licensing conditions limited the initial distribution. Meta's second-generation release responded to demand for a more broadly usable model family and incorporated additional training, evaluation, and alignment work. Meta introduced Llama 2 on July 18, 2023. The release comprised pretrained models with 7 billion, 13 billion, and 70 billion parameters, together with instruction-tuned chat versions. Meta reported that the models were trained on a larger public-data mixture than the original LLaMA and used a 4,096-token context length. The company also published a research paper, model card, responsible-use materials, and an acceptable-use policy. These materials described capabilities and limitations rather than presenting the models as reliable autonomous agents. A central feature of the launch was expanded availability. Meta offered downloads through its own request and distribution process and made Llama 2 available through a network of technology partners. Microsoft was presented as a preferred partner, with support through Azure and Windows-related developer channels. Other cloud and machine-learning providers subsequently offered hosted access, inference, or fine-tuning services. This partner strategy made Llama 2 usable by organizations that did not want to build all of the infrastructure required to run a model locally. Llama 2's product structure reflected two distinct developer needs. The pretrained checkpoints were general foundations for supervised fine-tuning and specialized applications. The Llama 2-Chat checkpoints were trained to follow conversational instructions, using supervised examples and reinforcement learning from human feedback. The chat variants were more convenient for assistant-like use, but Meta documented risks including hallucination, offensive or biased language, privacy concerns, prompt manipulation, and inappropriate reliance on generated answers. The model's licensing became an important part of its history. Meta permitted research and many commercial uses, but the Llama 2 license imposed conditions beyond the ordinary permissions associated with permissively licensed software. It included attribution obligations, requirements concerning redistribution and use of the Llama materials, and an additional provision directed at very large services. The accompanying acceptable-use policy prohibited certain applications. Consequently, the models were widely described as open or openly available, while some open-source organizations and developers argued that the license was not fully open source in the strict software-licensing sense. Llama 2 helped drive a rapid increase in experimentation with downloadable foundation models. Developers used it for chatbots, retrieval-augmented applications, summarization, coding tools, classification, education, and domain-specific fine-tuning. Its relatively accessible weights also supported local inference and a growing ecosystem of quantization, serving, evaluation, and customization tools. The 70B model targeted higher performance, while the smaller versions lowered the hardware threshold for experimentation and deployment. As Meta released later Llama generations, Llama 2 ceased to be the leading model in the company's own sequence, but it retained importance as a transition point. It demonstrated Meta's strategy of combining a high-profile consumer technology company with a broadly distributed model platform and helped normalize the idea that developers could download and operate a capable language model rather than access it only through a proprietary API.

• 2023

  Llama 2 announced

  [1]Introducing Llama 2ai.meta.com

  Meta released Llama 2 as a family of 7B, 13B, and 70B pretrained and chat-tuned language models for research and commercial use under a custom license.

• 2023

  Microsoft partnership expands distribution

  [4]Meta and Microsoft Introduce the Next Generation of Llamablogs.microsoft.com

  Microsoft announced Azure and Windows support for Llama 2, helping make the models available through mainstream cloud and developer channels.

• 2023

  Llama 2 paper and model documentation published

  [3]Llama 2: Open Foundation and Fine-Tuned Chat Modelsarxiv.org

  Meta and collaborators documented the model architecture, training approach, evaluations, safety work, and limitations in a research paper and associated model materials.

A broadly accessible, downloadable foundation-model family positioned for research, commercial development, customization, and self-hosted or cloud deployment.

• Pretrained language model · 2023

  Llama 2 7B

  The smallest standard Llama 2 model, intended to reduce deployment and experimentation requirements while providing a general-purpose language-model foundation. Developers could adapt it for specialized tasks or use it as the basis for applications where a larger model's quality and infrastructure cost were not justified.

• Pretrained language model · 2023

  Llama 2 13B

  A mid-sized Llama 2 checkpoint designed to balance model capability and deployment cost. It served as a general foundation for fine-tuning, text generation, question answering, summarization, and other natural-language applications, and was also offered in a chat-tuned form.

• Pretrained language model · 2023

  Llama 2 70B

  The largest principal Llama 2 model, created for stronger language performance and more demanding applications. Its scale generally required substantially more computing resources, but it was intended to deliver better results than the smaller checkpoints across a range of language tasks.

• Instruction-tuned language model · 2023

  Llama 2-Chat

  A family of dialogue-oriented Llama 2 checkpoints fine-tuned to follow instructions and conduct assistant-like conversations. Meta described the training as involving supervised fine-tuning and reinforcement learning from human feedback, while warning that the models could still generate inaccurate, biased, unsafe, or misleading responses.

• 2023

  Release model weights for broad developer accessProduct launch

  Background

  Meta sought to expand access to capable language models beyond a single hosted service and to build a developer ecosystem around its Llama research program.

  Action taken

  The company released Llama 2 pretrained and chat-tuned weights through Meta and partner distribution channels for research and many commercial uses.

  Aftermath

  The release accelerated experimentation with downloadable models and made Llama 2 a widely used base for hosted inference, local deployment, and fine-tuned derivatives.

  Source

• 2023

  Use a custom license rather than unrestricted open-source licensingStrategy

  Background

  Meta wanted to permit broad research and commercial use while retaining specific distribution, attribution, safety, and scale-related conditions.

  Action taken

  Meta published a Llama 2 Community License and a separate acceptable-use policy governing access, redistribution, and applications.

  Aftermath

  The approach enabled broad adoption but led to continuing debate about whether Llama 2 should be characterized as open source under strict licensing definitions.

  Source

• 2023

  Distribute through cloud and platform partnersStrategy

  Background

  Running large language models requires significant computing infrastructure, and partner distribution could reduce the barrier for enterprise and developer adoption.

  Action taken

  Meta worked with Microsoft and other ecosystem providers to make Llama 2 available through cloud, operating-system, and machine-learning platforms.

  Aftermath

  Llama 2 gained access to mainstream enterprise and developer workflows while remaining available for self-hosted use.

  Source

1. 1Introducing Llama 2ai.meta.com
2. 2Llama 2 Community Licenseai.meta.com
3. 3Llama 2: Open Foundation and Fine-Tuned Chat Modelsarxiv.org
4. 4Meta and Microsoft Introduce the Next Generation of Llamablogs.microsoft.com