This evergreen history article uses authoritative archives and official records. Exact dates are used when documented; gradual inventions and rollouts are described as periods rather than being assigned a misleading single birthday.
Quick facts
- Facebook created an artificial-intelligence research organization in 2013 that became widely known as FAIR.
- Meta contributed major open research and software, including PyTorch, computer-vision systems and language models.
- LLaMA 1 was announced in February 2023 primarily as a research foundation model.
- Llama 2 expanded access for research and many commercial uses, while Llama 3 and Llama 4 added stronger reasoning, multilingual and multimodal capabilities.
- Meta AI turned the models into a consumer assistant across WhatsApp, Instagram, Facebook, Messenger, web and a standalone app.
Facebook AI Research and open tools
Facebook formed an AI research laboratory in 2013 and recruited leading scientists including Yann LeCun. The organization worked on computer vision, translation, speech, recommendation and representation learning.
One of its most influential contributions was PyTorch, an open machine-learning framework that became widely used in research and production. This history established Meta as both a product company and a supplier of tools to the wider AI community.
The first LLaMA models
In February 2023 Meta announced LLaMA, short for Large Language Model Meta AI. The models were designed to give researchers capable systems at several sizes rather than requiring the largest possible model for every experiment.
The weights were initially distributed under research-focused conditions. After copies spread online, the event also demonstrated how difficult it is to control a powerful model once its files leave a company’s servers.
Llama 2 and wider commercial access
Meta and Microsoft announced Llama 2 in July 2023. The release included pretrained and chat-oriented models and allowed many commercial uses under Meta’s license.
The decision strengthened the open-weight ecosystem. Developers could run, fine-tune and deploy models on their own infrastructure, providing an alternative to AI available only through a closed cloud API.
Llama 3 and the Meta AI assistant
Llama 3 arrived in 2024 with improved instruction following and reasoning. Meta expanded the family through larger and smaller variants, multilingual support, vision models and on-device versions.
Meta AI used Llama technology inside WhatsApp, Instagram, Facebook and Messenger. The company integrated search, image generation, voice and creative tools into apps already used for communication and social media.
Llama 4 and native multimodality
In April 2025 Meta introduced Llama 4 Scout and Maverick as open-weight, natively multimodal mixture-of-experts models. A standalone Meta AI app followed, emphasizing voice interaction and a more personal assistant experience.
Meta’s strategy connected open model distribution with consumer products. External developers could build on Llama while Meta used related technology to improve its own assistant, advertising and content systems.
Open-weight does not mean unrestricted
Llama models are commonly called open source, but their custom licenses, usage conditions and incomplete training-data disclosure have led experts to prefer the term open-weight. The files can be downloaded, yet the development process is not open in every sense.
The approach offers control and customization but also makes misuse harder to prevent. Safety depends on model design, deployment choices, local safeguards and the responsibility of downstream developers.
Meta AI in 2026
By 2026, Meta AI had evolved from a chat feature into a more agentic assistant able to work across connected services and media. The Llama ecosystem continued to support research, startups and local deployment.
Meta’s history illustrates a distinctive AI strategy: publish powerful model weights widely, then use an enormous family of social products to distribute the assistant to consumers.
Timeline
| Year | Location | Event | Why it mattered |
|---|---|---|---|
| 2013 | Menlo Park, California and global labs | Facebook AI Research is established | Created Meta’s long-term research foundation. |
| 2016–2017 | Global developer community | PyTorch is released and open-sourced | Became a major framework for deep-learning research. |
| February 2023 | United States | LLaMA 1 is announced | Introduced Meta’s modern family of foundation language models. |
| July 2023 | Global | Llama 2 expands research and commercial access | Strengthened the open-weight AI ecosystem. |
| April–July 2024 | Global | Llama 3 and Llama 3.1 launch | Improved capability and introduced a frontier-scale 405B model. |
| April 2025 | Global | Llama 4 and the standalone Meta AI app arrive | Combined multimodal open-weight models with a consumer assistant platform. |
Frequently asked questions
What does Llama stand for?
The original name LLaMA expanded to Large Language Model Meta AI.
Is Llama open source?
The model weights are broadly available under Meta licenses, but because the licenses and development process are not fully open by every definition, many researchers call Llama open-weight.
Where can people use Meta AI?
Meta has offered the assistant across WhatsApp, Instagram, Facebook, Messenger, the web and a standalone app, with availability varying by country.
Why does Meta release model weights?
Meta argues that broader access supports research and innovation, while the ecosystem also increases adoption of Meta’s technology and tools.