LLaMA

LLaMA, commonly branded as Llama, is a family of large language models developed and released by Meta AI beginning in February 2023. The name is a backronym for “Large Language Model Meta AI.” Rather than being a single application, Llama is a model platform spanning pretrained foundation models, instruction-tuned conversational models, coding variants, multimodal systems, and integrations such as Meta AI. The family is designed to support a broad range of uses, from academic experimentation and local inference to enterprise software, developer tools, chat assistants, and generative AI products. The first Llama release emphasized efficiency and accessibility. Meta published the model’s research paper and inference code and offered model-weight access through an application process, initially restricting availability largely to researchers and other qualified organizations under a non-commercial license. Llama 1 was released in several sizes, allowing researchers to run models on more modest hardware than the largest competing systems. Its reported benchmark performance, particularly for the 13-billion-parameter model, helped establish Llama as an important alternative to much larger proprietary systems. The model weights were nevertheless leaked online shortly after launch, accelerating independent experimentation and creating a dispute over unauthorized distribution. Llama 2, announced in July 2023 in partnership with Microsoft, broadened the project’s commercial and developer reach. It was offered in 7-billion, 13-billion, and 70-billion-parameter versions, with both foundation and chat-oriented models. Meta’s license permitted many commercial uses, although its acceptable-use restrictions meant that the release did not meet the Open Source Initiative’s formal definition of open source. This licensing debate became a recurring feature of Llama’s positioning: Meta presented the family as open and widely accessible, while critics distinguished between publicly available weights and fully open-source software and data. The Llama ecosystem expanded through specialized variants. Code Llama adapted Llama 2 for programming, code completion, and code-related instruction following. Llama 3, released in April 2024, introduced 8-billion- and 70-billion-parameter models trained on a substantially larger corpus, with improvements in reasoning, coding, instruction following, and general language performance. Llama 3.1 subsequently added an especially large 405-billion-parameter model alongside updated 8-billion and 70-billion versions. These releases reinforced Meta’s strategy of publishing capable models for external deployment while also using them internally. Llama 4, released in 2025, marked a major architectural and product shift. Its Scout and Maverick models used mixture-of-experts designs, in which only selected expert components are activated for each token, and accepted both text and image input while producing text output. The release also targeted multilingual use and much longer context windows. Meta described Behemoth as a larger model under development rather than a generally available release. Alongside Llama 3 and later generations, Meta promoted Meta AI, an assistant available through a dedicated website and Meta services including Facebook and WhatsApp. Llama’s strategic importance lies in the combination of model capability, downloadable weights, multiple parameter scales, and a large surrounding ecosystem. Developers can adapt or deploy the models through cloud services, hardware vendors, open machine-learning tools, and self-hosted infrastructure. At the same time, the family has faced criticism and legal scrutiny involving licensing, training-data provenance, safety, benchmark practices, and the tension between broad availability and responsible-use controls. Llama remains a central component of Meta’s artificial-intelligence strategy and a prominent competitor to proprietary model families from companies such…

Llama emerged from Meta AI’s response to the rapid progress of large language models and the public impact of systems such as GPT-3 and ChatGPT. Meta’s research direction emphasized improving capability through efficient architectures, high-quality training data, and a range of parameter sizes rather than relying only on ever-larger models. This made the family suitable for researchers and organizations that lacked access to the largest proprietary systems. Meta announced the first Llama model on February 24, 2023. The release included a research paper, inference code, and several foundation-model sizes. The models were trained primarily on publicly available material and used a decoder-only transformer architecture with design choices including SwiGLU activations, rotary positional embeddings, and RMSNorm. Unlike later releases, the original distribution focused on pretrained foundation models rather than a broadly available chat product. Access to weights was managed through an application process, and the associated license limited use, particularly for commercial purposes. Within days, copies of the initial weights appeared through torrent and mirror links. The incident materially changed the project’s reach: independent developers and researchers could experiment with Llama without waiting for individual approval, but Meta treated the redistribution as unauthorized and pursued takedown requests against repositories and downloading tools. The episode also fueled debate over whether controlled access was compatible with the research community’s expectations and whether relatively compact models could accelerate useful or harmful applications. Llama 2 represented a deliberate expansion. Announced on July 18, 2023, with Microsoft, it provided 7B, 13B, and 70B versions, plus chat-tuned models. Meta used more training data than for the first generation and incorporated supervised fine-tuning and reinforcement learning from human feedback for the conversational versions. The release made model weights available for many commercial use cases, creating a much larger developer ecosystem. However, the license included an acceptable-use policy and other conditions, leading the Open Source Initiative and other commentators to reject Meta’s use of “open source” in the strict technical sense. The family then diversified. Code Llama, released in August 2023 and expanded with additional variants in 2024, applied Llama 2 technology to programming tasks. Meta also used Llama as a foundation for its own assistant strategy. With Llama 3, released on April 18, 2024, Meta delivered 8B and 70B models trained on roughly 15 trillion tokens, including mainly English data and material in numerous other languages. Instruction-tuned versions were trained with human-annotated examples. The release coincided with a broader rollout of Meta AI through Meta’s website and social platforms. Llama 3.1, released on July 23, 2024, added an updated 405B model and strengthened the family’s position as a platform for external and internal applications. Llama 4, released in April 2025, changed the technical direction through mixture-of-experts models. Scout and Maverick activated only portions of their expert networks for each input, supporting larger total parameter counts while controlling active computation. They also accepted images as well as text and targeted multilingual operation. Scout emphasized an exceptionally long context window, while Maverick offered a larger mixture-of-experts configuration. Meta announced Behemoth as a substantially larger system still in training rather than as a generally released model. Throughout its development, Llama has occupied an unusual position in the AI market. It is not merely a consumer brand, because the principal products are model families and associated weights, licenses, tools, and services. Nor is it fully open in the same way as projects that publish unrestricted code, data, and weights. Its influence comes from the ability of cloud providers, hardware companies, researchers, startups, and individual developers to build on published model artifacts. Its controversies have included the first-model leak, disputes over licensing terminology, questions about training-data rights, safety concerns, and criticism of benchmark claims involving an experimental Llama 4 version. Despite these issues, Meta continues to use Llama as a core technological and strategic foundation for its generative-AI products.

• 2025

  Llama 4 generation released

  [7]The Llama 4 herd: Advancing multimodal intelligenceai.meta.com

  Released in April 2025, Scout and Maverick introduced mixture-of-experts and multimodal model designs; the supplied reference material dates the releases to April 5, 2025.

• 2025

  Llama 4 Scout and Maverick

  [1]Llama (language model)en.wikipedia.org

  The supplied reference material dates the Llama 4 Scout and Maverick releases to April 5, 2025.

• 2024

  Llama 3.1 released

  [5]Introducing Meta AI with Llama 3ai.meta.com

  Released on April 18, 2024, the third generation launched with 8B and 70B models trained on roughly 15 trillion tokens and coincided with a broader rollout of Meta AI. Released on July 23, 2024, the updated family added a 405B model and enhanced the smaller 8B and 70B versions.

• 2024

  Llama 3 and Meta AI

  [1]Llama (language model)en.wikipedia.org

  Meta released Llama 3 and introduced Meta AI as a user-facing assistant built on Llama technology.

• 2023

  First Llama model announced

  [2]Introducing LLaMA: A foundational, 65-billion-parameter language modelai.meta.com

  Meta AI introduced the first Llama family model with multiple parameter sizes, a research paper, and publicly released inference code; access to weights was initially controlled. Meta AI introduced the first Llama family model on February 24, 2023, with multiple parameter sizes, a research paper, and publicly released inference code; access to weights was initially controlled through an application process under a non-commercial research license.

• 2023

  Llama weights leaked

  [1]Llama (language model)en.wikipedia.org

  Unauthorized copies of the first model’s weights spread through torrent and mirror links, including BitTorrent, prompting Meta takedown requests against repositories and downloading tools.

• 2023

  Llama 2 launched with Microsoft

  [3]Llama 2ai.meta.com

  Announced on July 18, 2023, the second generation added 7B, 13B, and 70B versions, chat-tuned models, broader access, and licensing allowing many commercial deployments.

• 2023

  Code Llama released

  [4]Introducing Code Llamaai.meta.com

  Meta introduced a programming-focused fine-tune of Llama 2 in several sizes in August 2023, later expanded with additional variants in 2024.

• 2023

  First Llama announcement

  [1]Llama (language model)en.wikipedia.org

  Meta AI announced the first Llama model on February 24, 2023, alongside technical documentation and inference code.

• 2023

  Unauthorized distribution of initial weights

  [1]Llama (language model)en.wikipedia.org

  Unauthorized copies of the first model weights were circulated online, including through BitTorrent, after the initial application-based release.

• 2023

  Llama 2 broadens availability

  [1]Llama (language model)en.wikipedia.org

  Llama 2 expanded access beyond the original research setting and permitted certain commercial uses under its applicable license.

A broadly accessible, high-capability model family positioned between proprietary AI services and traditional open-source machine-learning ecosystems. Meta emphasizes downloadable weights, model variety, developer adoption, and integration across its products, while licensing and acceptable-use restrictions distinguish Llama from fully open-source projects. Meta’s open-weight and commercially usable large-language-model platform, spanning research models, instruction-tuned systems, developer deployments, and consumer AI experiences.

• Foundation language model · 2023

  Llama 1

  The original 2023 Llama family consisted of pretrained, decoder-only transformer models in multiple sizes. It was trained on publicly available material and designed to provide strong language performance with comparatively manageable inference requirements. The initial release was aimed primarily at researchers and used a controlled-access, non-commercial licensing arrangement. It did not initially include a generally distributed instruction-tuned chat family. Llama 3 introduced 8B and 70B models trained on a substantially expanded token corpus, with improvements in general language understanding, instruction following, coding, and reasoning. Llama 3.1 extended the line with updated 8B and 70B systems and a 405B flagship. These models became important foundations for Meta AI and for third-party applications built through cloud, hardware, and local-inference ecosystems. The original Llama release, announced in February 2023, was primarily a foundation model for research and downstream experimentation. Meta published technical material and inference code, while access to the weights was initially managed through an application process and restricted by a non-commercial research license.

• Foundation and chat language models · 2023

  Llama 2

  Llama 2 expanded the family into 7B, 13B, and 70B versions and paired pretrained models with chat-tuned variants. Meta used a larger training corpus and human-feedback methods to improve dialogue safety and usefulness. Its license enabled many commercial applications and helped establish Llama as a widely adopted base for hosted, embedded, and self-managed AI systems, although its restrictions prevent it from being unambiguously classified as open source. Llama 2 expanded the family’s availability beyond the initial research-only distribution model. Meta released foundation and instruction-tuned variants, making the generation useful for both model adaptation and direct conversational or instruction-following applications. Its licensing permitted certain commercial uses, subject to the specific terms of the release.

• Code generation model · 2023

  Code Llama

  Code Llama is a Llama 2-derived family specialized for software development. It was trained on code-oriented data and offered foundation, instruction-following, and Python-focused variants. The models support code completion, generation, explanation, debugging, and related programming workflows. Releases covered several parameter scales, allowing use in both larger hosted environments and more resource-constrained developer setups.

• Multimodal mixture-of-experts model · 2025

  Llama 4 Scout

  Scout is a Llama 4 model using a mixture-of-experts architecture, with a relatively small number of active parameters per token compared with its total parameter count. It accepts text and image inputs, produces text, supports multiple languages, and is designed around an exceptionally long context window. Meta released both foundation and instruction-tuned forms. The supplied reference material dates the release to April 5, 2025. Llama 4 Scout is one of the Llama 4 releases identified in the supplied reference material. It was dated April 5, 2025 and represents a later-generation model in Meta’s continuing family of differently sized and configured language models.

• Multimodal mixture-of-experts model · 2025

  Llama 4 Maverick

  Maverick is a larger Llama 4 mixture-of-experts model with multimodal and multilingual capabilities. Its design activates a subset of expert networks for each token, seeking a balance between broad model capacity and practical inference cost. Meta released foundation and instruction-tuned versions and positioned Maverick for demanding general-purpose and conversational workloads. The supplied reference material dates the release to April 5, 2025. Llama 4 Maverick is a later-generation Llama release identified in the supplied reference material as dating from April 5, 2025. It forms part of Meta’s strategy of offering multiple model configurations for different capability and deployment requirements.

• AI assistant built on Llama · 2024

  Meta AI

  Meta AI is Meta’s consumer-facing assistant strategy associated with the Llama family. It is available through a dedicated web presence and Meta platforms including Facebook and WhatsApp. The assistant translates the underlying model technology into conversational search, writing, information, and creative-use experiences, while also serving as a major channel through which Meta distributes Llama-based capabilities to consumers. Meta AI is a user-facing assistant built on Llama technology. It was introduced alongside the Llama 3 era and made available through a dedicated website and Meta platforms including Facebook and WhatsApp, extending Llama from a developer model family into a consumer service.

• Large language model family · 2024

  Llama 3

  Llama 3 continued Meta’s general-purpose model series and supported the company’s expansion into consumer-facing AI. Its release was associated with Meta AI, an assistant made available through a dedicated website and Meta services such as Facebook and WhatsApp.

• 2025

  Dispute over Llama 4 benchmark presentation

  Critics objected that Meta highlighted results from an experimental conversational version of Llama 4 Maverick rather than the publicly released model. LMArena said Meta’s presentation did not match its expectations for disclosure and indicated that it would change policies; Meta also denied allegations that it trained on benchmark test sets.

  Source

• 2023

  Unauthorized distribution of Llama 1 weights

  The first Llama weights were redistributed online without Meta’s authorization, including through BitTorrent. Meta sought removal of linked repositories and downloading scripts, while researchers and developers debated the benefits and risks of the wider access.

  Source

• 2023

  Copyright and training-data litigation

  Authors sued Meta over allegations concerning the use of copyrighted books and other material in training language models. The allegations have been contested and the legal issues concern copyright, dataset construction, and disclosure of training practices.

  Source

Read more

• 2025

  Adopt mixture-of-experts and multimodal architecture in Llama 4Generation change

  Background

  Meta’s later-generation models needed greater capacity, long-context handling, and image understanding without activating every parameter for every token.

  Action taken

  Llama 4 Scout and Maverick used mixture-of-experts designs and accepted both text and image inputs.

  Aftermath

  The family expanded beyond text-only language modeling toward multimodal and multilingual platform capabilities.

  Source

• 2024

  Integrate Llama into Meta AIProduct launch

  Background

  Meta sought to turn its published model family into a consumer-facing assistant and distribution channel. Meta expanded the role of Llama from downloadable or research-oriented models into a consumer-facing assistant ecosystem.

  Action taken

  Meta rolled out Meta AI through a dedicated website and services including Facebook and WhatsApp, using Llama as a principal technology foundation.

  Aftermath

  Llama became both an external developer platform and an internal foundation for Meta’s consumer AI products. Llama gained a direct consumer distribution channel in addition to its developer and research uses.

  Source

• 2023

  Move from controlled research access toward broad developer availabilityStrategy

  Background

  The first Llama release used an application process and restrictive licensing, while demand for leaked weights demonstrated strong interest from independent developers and researchers.

  Action taken

  Meta released Llama 2 weights more broadly and permitted many commercial uses, while retaining an acceptable-use policy and other license conditions.

  Aftermath

  The decision helped create a large Llama ecosystem but also intensified debate over whether the models should be called open source.

  Source

• 2023

  Controlled initial release of Llama weightsStrategy

  Background

  Meta introduced Llama as a research-oriented foundation model and initially limited access to approved applicants under a non-commercial license.

  Action taken

  The company published inference code and technical information while distributing the model weights through an application-based process.

  Aftermath

  Unauthorized copies later circulated online, increasing the model’s availability beyond Meta’s intended channel.

  Source

• 2023

  Broader licensing and distribution with Llama 2Strategy

  Background

  Meta sought to make the model family more useful outside academic research and to support a wider developer and commercial ecosystem.

  Action taken

  With Llama 2, Meta expanded access and used licensing that permitted certain commercial applications, while releasing instruction-tuned variants alongside foundation models.

  Aftermath

  Llama became more accessible to organizations building applications, fine-tuning models, and deploying language-model systems.

  Source

1. 1Llama (language model)en.wikipedia.org
2. 2Introducing LLaMA: A foundational, 65-billion-parameter language modelai.meta.com
3. 3Llama 2ai.meta.com
4. 4Introducing Code Llamaai.meta.com
5. 5Introducing Meta AI with Llama 3ai.meta.com
6. 6The Llama 3.1 Herd of Modelsai.meta.com
7. 7The Llama 4 herd: Advancing multimodal intelligenceai.meta.com