Meta AI

Meta AI is the artificial-intelligence research and product organization of Meta Platforms, Inc. It originated in 2013 as Facebook Artificial Intelligence Research, commonly known as FAIR, and initially focused on advancing foundational AI science through academic collaboration, open publication, and the development of systems for language, vision, translation, recommendation, and machine learning. After Facebook, Inc. became Meta Platforms in 2021, FAIR and related research programs were progressively presented under the broader AI at Meta and Meta AI identity. The organization has operated as both a research institution and a commercial technology platform. Its research portfolio has included self-supervised learning, generative adversarial networks, computer vision, document classification, machine translation, natural-language generation, speech and multimodal interaction. Meta AI’s work has also produced or supported widely used open-source and source-available tools, including fastText, Torch modules, PyTorch, and the Llama family of large language models. These releases helped establish Meta as an important participant in the global AI infrastructure and developer ecosystem rather than merely an internal consumer-product organization. The Meta AI name is also used for Meta’s consumer-facing virtual assistant. First introduced publicly in 2023 and expanded substantially in 2024, the assistant is integrated into WhatsApp, Instagram, Facebook, Messenger, Meta’s web experience at meta.ai, Ray-Ban Meta smart glasses, and Meta Quest devices. Depending on the surface and market, it can answer questions, search for information, summarize or explain material, generate and edit images, participate in voice conversations, interpret photographs, translate text, and assist with everyday planning. Meta positions the assistant as a free, widely distributed alternative to standalone chatbot services, using the reach and social context of Meta’s existing platforms as a principal competitive advantage. The product’s underlying technology has evolved with Meta’s model releases. Meta AI was initially associated with Llama-based systems, including Llama 2 and Llama 3. The Llama 3 launch in 2024 supported a major international rollout and added faster image generation, web search, and broader availability across Meta’s applications. Later in 2024, Meta added voice interaction, photo understanding, and multimodal capabilities. In April 2025, Meta launched a standalone Meta AI application built with Llama 4, adding personalized conversations, memory features, a Discover feed, voice-first interaction, and continuity between the app, the website, and AI-enabled glasses. In 2026, Meta Superintelligence Labs introduced Muse Spark, which Meta said would power the Meta AI app and website and gradually extend across WhatsApp, Instagram, Facebook, Messenger, Threads, and AI glasses. Meta AI’s positioning combines open model distribution, consumer-scale deployment, and personal assistance. Meta has emphasized releasing models, tools, safety resources, and research to external developers, although the licensing terms of successive Llama versions have varied and are not identical to traditional permissive open-source licenses. At the product level, Meta emphasizes convenience, personalization, multimodality, and integration with social relationships and user context. The approach gives Meta access to enormous distribution through its existing services, but it also creates questions about privacy, advertising personalization, information quality, the use of news content, copyright in training data, and the safety of AI systems embedded in social networks and wearable devices. In 2025, Meta reorganized much of its advanced AI work under Meta Superintelligence Labs, bringing together FAIR, large-language-model teams, and other research and product groups. This reorganization placed Meta AI within a more centralized strategy aimed at developing persona…

Meta AI began in 2013 as Facebook Artificial Intelligence Research, or FAIR. The organization was established to conduct fundamental research in artificial intelligence while maintaining strong links with universities and the wider research community. Yann LeCun became its initial scientific leader, and the early group worked on machine learning, computer vision, language, robotics, and methods that could learn from large amounts of data with limited human labeling. During its first years, FAIR expanded through a network of international laboratories and academic partnerships. The group opened or strengthened research activity in locations including Menlo Park, New York, Paris, London, Seattle, Pittsburgh, Tel Aviv, and Montreal. Its research agenda included self-supervised learning, generative adversarial networks, translation, language understanding, document classification, visual recognition, and recommendation technologies. In 2016, FAIR joined other major technology companies and academic organizations in the Partnership on AI, reflecting Facebook’s public emphasis on collaborative research and responsible development. A significant part of FAIR’s influence came through released software and research artifacts. In 2016, the organization released fastText, a library designed to make text classification and word representation learning efficient, including through character-level information. In 2017, Facebook researchers released Torch-related deep-learning modules and PyTorch. PyTorch later became one of the most widely used frameworks for deep-learning research and production systems, supporting work in universities, startups, and large technology companies. These releases helped make FAIR’s work relevant beyond Facebook’s own products. LeCun led FAIR until 2018, when Jérôme Pesenti succeeded him as director. The leadership transition occurred as Facebook was increasing investment in applied machine learning, platform integrity, recommendation systems, and commercial AI. FAIR continued to publish research and cooperate with universities, while its technologies became increasingly connected to content ranking, translation, image understanding, moderation, and other systems used across Facebook’s services. FAIR also attracted public attention in 2017 after reports claimed that Facebook had shut down chatbots because they had invented a secret language. The underlying research was more limited: the bots had been trained to negotiate with one another, and their abbreviated exchanges were not intelligible to people. Facebook said the experiment was stopped or adjusted because researchers had met the immediate research objective and wanted to continue under more controlled conditions. The episode became an early example of how AI research could be misunderstood when simplified through sensational reporting. In November 2021, Facebook, Inc. changed its corporate name to Meta Platforms, Inc. The research organization subsequently operated under the Meta AI and Fundamental AI Research identities. The rebranding connected AI work to Meta’s broader ambition in augmented and virtual reality, including computer vision, embodied interaction, speech, and wearable computing. Meta AI’s research scope therefore extended beyond conventional software models to systems intended to interpret the physical world and assist people through devices such as smart glasses and virtual-reality headsets. In November 2022, Meta AI released Galactica, a language model and public demonstration designed to generate scientific text. The demonstration was removed after only a few days because users produced and observed racist, offensive, and factually unreliable outputs presented in a scientific style. Galactica became an influential case in discussions about hallucination, scientific misinformation, model evaluation, and the risks of publicly releasing a generative system before its limitations are well understood. In February 2023, Meta released the first Llama large language model. The original release was primarily aimed at researchers and was distributed under controlled terms, but unauthorized copies soon circulated online. Later Llama versions were made available more broadly and supported commercial and developer use under their respective licenses. Llama 2, Llama 3, and Llama 4 established Meta as one of the most visible providers of openly distributed or open-weight foundation models. Meta used the models internally while encouraging cloud providers, application developers, and researchers to build on them. Meta introduced its consumer Meta AI assistant in 2023. The assistant brought Meta’s foundation-model research directly into WhatsApp, Instagram, Facebook, Messenger, and later the web experience at meta.ai. Its capabilities expanded from text conversation to web search, image generation, voice interaction, photo understanding, translation, and contextual assistance. The April 2024 Llama 3 rollout marked a major international expansion, while the same period saw Meta AI integrated into Ray-Ban Meta smart glasses with camera-based vision functions. Meta later announced Meta AI for Quest devices, extending the assistant into mixed-reality environments. In 2024, Meta AI also highlighted Seamless Communication, a family of multilingual speech and translation models. The system combined speech-to-text, text-to-speech, translation, streaming, and expressive voice preservation. Meta released the models and supporting resources publicly, positioning them as tools for reducing language barriers while acknowledging safety concerns such as toxicity, hallucinated translations, and misuse of synthetic voices. In April 2025, Meta launched a standalone Meta AI application built with Llama 4. The app was designed to give the assistant a persistent consumer destination outside Meta’s existing social applications. It added voice-first interaction, personalization, memory, image creation and editing, a Discover feed, and links to the web experience and AI glasses. Meta subsequently stated that more than one billion people used its AI products each month, although that figure referred to Meta’s broader AI ecosystem rather than necessarily to unique users of the standalone assistant. Meta reorganized its advanced AI activities in 2025 under Meta Superintelligence Labs. The new structure brought together FAIR, large-language-model teams, and related research and product groups. The reorganization reflected Meta’s shift from treating AI primarily as a collection of research projects and product features toward a central corporate strategy focused on personal AI, advanced reasoning, specialized hardware, and eventual superintelligence. In October 2025, Meta also announced that interactions with its AI features would be used to personalize content and advertising recommendations, prompting renewed privacy and data-governance concerns. In April 2026, Meta Superintelligence Labs announced Muse Spark, the first model in its Muse series. Meta said the model powered the Meta AI application and meta.ai and would roll out progressively across WhatsApp, Instagram, Facebook, Messenger, Threads, and AI glasses. The announcement represented a further transition from the Llama-branded foundation-model era toward a more integrated assistant stack designed around Meta’s products, user context, multimodal interaction, and personalized AI. Meta AI therefore remains both a research lineage descended from FAIR and a major consumer-facing AI brand embedded throughout Meta’s global technology ecosystem.

1. 2026Muse Spark is introduced

   Meta Superintelligence Labs announces Muse Spark as the first model in a new series intended to power Meta AI across apps, the web, and AI glasses.

   Source
2. 2025The standalone Meta AI app launches

   Meta releases a dedicated Meta AI application with voice interaction, personalization, image tools, memory, and a Discover feed.

   Source
3. 2025Meta Superintelligence Labs is formed

   Meta consolidates FAIR and other advanced AI research and product teams under Meta Superintelligence Labs.

   Source
4. 2025AI interactions become a personalization signal

   Meta announces that interactions with its AI features will influence content and advertising recommendations from December 16, 2025.

   Source
5. 2024Meta AI expands internationally with Llama 3

   Meta expands the assistant to additional countries and adds web access, image generation, and broader integration across Facebook, Instagram, WhatsApp, Messenger, and meta.ai.

   Source
6. 2024Meta AI gains vision on Ray-Ban Meta glasses

   Meta rolls out camera-aware Meta AI features on supported Ray-Ban Meta smart glasses in the United States and Canada.

   Source
7. 2024Seamless Communication models are released

   Meta publishes a family of multilingual speech and translation models designed for more natural, expressive, and lower-latency cross-language communication.

   Source
8. 2023Llama and the Meta AI assistant emerge as major product pillars

   Meta releases the first Llama model and introduces its consumer Meta AI assistant, connecting foundation-model research to its social and messaging products.

   Source
9. 2022Galactica public demonstration is withdrawn

   Meta removes the Galactica demonstration after criticism over inaccurate, offensive, and misleading scientific-looking outputs.

   Source
10. 2021Facebook becomes Meta

    The corporate rebranding to Meta Platforms leads FAIR and related programs to be presented under the Meta AI and Fundamental AI Research identities.

    Source
11. 2018FAIR leadership changes

    Jérôme Pesenti succeeds Yann LeCun as director of Facebook AI Research.

    Source
12. 2017PyTorch is released

    FAIR releases PyTorch and related Torch modules, helping establish a widely adopted deep-learning research and production framework.

    Source
13. 2016fastText is released

    FAIR releases fastText, an efficient open-source library for text classification and word representation learning.

    Source
14. 2016Facebook joins the Partnership on AI

    Facebook collaborates with Google, Amazon, IBM, and Microsoft in establishing the Partnership on AI to Benefit People and Society.

    Source
15. 2015FAIR opens a Paris research center

    Facebook expands its international AI research network with a Paris laboratory focused on collaboration with European researchers.

    Source
16. 2013Facebook AI Research is founded

    Facebook establishes Facebook AI Research, later known as FAIR and then Meta AI, to conduct fundamental artificial-intelligence research.

    Source

A globally distributed, consumer-scale AI platform that combines open model research with an assistant embedded in Meta’s social, messaging, wearable, and immersive-computing products.

• Meta AI assistantConversational AI assistant2023

  Meta’s consumer-facing conversational assistant, available through WhatsApp, Instagram, Facebook, Messenger, meta.ai, the Meta AI app, Ray-Ban Meta glasses, and selected Quest devices. Depending on the product and market, it supports text and voice conversation, web search, question answering, planning, translation, image generation, photo understanding, image editing, and contextual assistance.

• Meta AI appStandalone AI application2025

  A dedicated mobile application launched in 2025 to provide a persistent home for Meta AI outside Meta’s social and messaging applications. It supports text and voice interaction, personalized responses, memory, image generation and editing, a Discover feed, and continuity with meta.ai and AI glasses.

• LlamaLarge language model family2023

  Meta’s family of large language models, beginning with the original Llama release in 2023 and continuing through Llama 2, Llama 3, and Llama 4. Meta has distributed successive versions to researchers, developers, cloud providers, and organizations under version-specific terms, making Llama a central part of Meta’s open-model strategy and the technical foundation of early Meta AI assistant releases.

• PyTorchMachine-learning framework2017

  An open-source deep-learning framework released by FAIR in 2017 with roots in Facebook’s Torch work. PyTorch became a major tool for AI research and production development and helped extend Meta AI’s influence beyond Meta’s own applications.

• fastTextNatural-language-processing library2016

  An open-source library released by FAIR in 2016 for efficient text classification and learning word representations. Its use of subword and character-level information made it useful for handling rare words, spelling variation, and multilingual text with relatively modest computational requirements.

• Seamless CommunicationSpeech translation and multilingual AI2023

  A family of publicly released models including SeamlessM4T v2, SeamlessStreaming, SeamlessExpressive, and the unified Seamless system. The models target speech and text translation, low-latency communication, preservation of vocal style and prosody, and more natural multilingual interaction.

• GalacticaScientific-text language model2022

  A large language model and public demonstration designed to generate scientific text and manipulate scientific knowledge. The public demonstration was withdrawn after users identified inaccurate, offensive, and misleading outputs, making it an important example of the risks of deploying generative models in high-trust domains.

• MTIAAI accelerator hardware2022

  Meta’s internally designed AI accelerator line, developed to support workloads such as content recommendation and other large-scale platform systems. MTIA formed part of Meta’s effort to supplement or replace general-purpose compute with specialized infrastructure for AI inference and training.

• Muse SparkLarge language model2026

  The first model announced by Meta Superintelligence Labs’ Muse series. Meta describes Muse Spark as a small and fast model capable of multimodal tasks and complex reasoning, powering the Meta AI app and meta.ai before a broader rollout across Meta’s social, messaging, and wearable products.

• 2026Muse Spark rollout

  Global

  Meta introduced Muse Spark as the first model in a new series and promoted it as a faster, more capable foundation for Meta AI across the app, website, social platforms, messaging services, Threads, and AI glasses.

  Outcome. Muse Spark marked Meta’s shift toward a new model family developed by Meta Superintelligence Labs and a more unified assistant experience across Meta products.

  Source
• 2025The Meta AI app

  United States · Canada · Australia · New Zealand · Global expansion

  Meta launched a standalone app centered on voice conversations, personalization, memory, image creation, and a Discover feed. The product was presented as a more personal assistant connected to the company’s broader ecosystem.

  Outcome. The standalone app gave Meta a direct consumer destination for AI while retaining distribution through WhatsApp, Instagram, Facebook, Messenger, the web, and smart glasses.

  Source
• 2025Personal superintelligence for everyone

  Global

  Mark Zuckerberg and Meta framed the company’s AI strategy around personal superintelligence: advanced AI placed directly in people’s hands and adapted to individual goals, relationships, and daily activities.

  Outcome. The message provided the strategic narrative for consolidating Meta’s AI research, models, products, infrastructure, and devices under a more centralized advanced-AI program.

  Source
• 2024Meta AI international rollout

  United States · Australia · Canada · Ghana · Jamaica · Malawi · New Zealand · Nigeria

  Meta promoted Meta AI as a free assistant available across the company’s major social and messaging platforms, using Llama 3 to support question answering, real-time web information, image creation, and everyday planning.

  Outcome. The assistant expanded from a primarily U.S.-focused launch into a global product rollout and became a central consumer-facing component of Meta’s AI strategy.

  Source
• 2024Meta AI with Vision for Ray-Ban Meta

  United States · Canada

  Meta marketed camera-aware AI on Ray-Ban Meta smart glasses, emphasizing hands-free assistance, visual question answering, translation, and information about objects and surroundings.

  Outcome. The campaign helped position Meta AI as a wearable and multimodal product rather than only a text chatbot.

  Source

• 2026Replace the early Llama-based assistant stack with Muse SparkGeneration change

  After forming Meta Superintelligence Labs, Meta sought a new model family designed specifically for its products and personal-AI ambitions.

  What changed. Meta announced Muse Spark as the first Muse-series model and said it powered the Meta AI app and meta.ai, with rollout planned across Meta’s messaging, social, and wearable products.

  Aftermath. The announcement marked a transition from Llama as the primary public identity behind Meta AI toward a more integrated model-and-product architecture controlled by Meta Superintelligence Labs.

  Source
• 2025Launch the standalone Meta AI app using Llama 4Product launch

  Meta wanted Meta AI to have a direct consumer destination rather than relying exclusively on entry points inside existing social and messaging products.

  What changed. Meta launched a standalone application with voice interaction, personalization, memory, image tools, a Discover feed, and links to meta.ai and AI glasses.

  Aftermath. The app strengthened Meta’s direct competition with standalone AI assistants and made personal context, voice, and cross-device continuity central to the brand.

  Source
• 2025Consolidate advanced AI efforts under Meta Superintelligence LabsStrategy

  Meta’s AI work had become distributed across FAIR, language-model development, product teams, infrastructure groups, and hardware programs.

  What changed. Meta created Meta Superintelligence Labs to bring together advanced research, large-language-model development, and related AI product work.

  Aftermath. Meta AI became part of a more centralized advanced-AI strategy focused on personal AI, frontier models, specialized computing, and eventual superintelligence.

  Source
• 2025Use AI interactions for content and advertising personalizationStrategy

  Meta sought to use conversational AI as another source of behavioral signals for improving relevance across its platforms.

  What changed. Meta announced that voice and text interactions with its AI features would be used to personalize content and advertising recommendations beginning December 16, 2025.

  Aftermath. The decision intensified debate about whether AI conversations should be treated like ordinary engagement data, particularly when conversations may reveal sensitive interests or intentions.

  Source
• 2024Rebuild Meta AI around Llama 3Generation change

  Meta sought to improve the assistant’s quality and broaden its international reach while using the distribution of Facebook, Instagram, WhatsApp, and Messenger.

  What changed. Meta connected Meta AI to Llama 3, added web search and faster image generation, and began expanding availability to more than a dozen countries outside the United States.

  Aftermath. Meta AI shifted from a limited assistant experiment toward a global, multimodal consumer platform embedded across Meta’s products.

  Source
• 2023Launch the Meta AI consumer assistantProduct launch

  Meta had developed large language models through FAIR and was seeking a consumer product that could compete with rapidly growing standalone chatbot services.

  What changed. Meta introduced a conversational Meta AI assistant based on its Llama models and integrated it into selected social and messaging experiences.

  Aftermath. The assistant became the principal consumer brand through which Meta delivered generative AI and later expanded to the web, image generation, voice, smart glasses, and Quest devices.

  Source
• 2022Release and withdrawal of GalacticaProduct launch

  Meta AI sought to demonstrate a language model specialized for scientific text and knowledge.

  What changed. Meta released Galactica publicly on November 15, 2022, then withdrew the demonstration approximately three days later after widespread criticism of inaccurate and offensive outputs.

  Aftermath. The episode became a cautionary example in discussions about model evaluation, hallucination, scientific misinformation, and staged deployment of generative AI.

  Source
• 2021Rebrand FAIR-related work under Meta AIStrategy

  Facebook, Inc. changed its corporate identity to Meta Platforms as it expanded its ambitions in artificial intelligence, augmented reality, virtual reality, and the metaverse.

  What changed. The company increasingly presented FAIR and related programs as Meta AI or Fundamental AI Research, connecting foundational research to a wider portfolio of consumer products and devices.

  Aftermath. Meta AI became a broader organizational and product identity spanning research, foundation models, conversational assistants, computer vision, translation, and hardware experiences.

  Source

• 2025Copyright lawsuits over AI training materialsControversy

  Authors and publishers brought litigation alleging that Meta used copyrighted books and other works, including material obtained from shadow-library sources such as LibGen, in training language models. Meta disputed liability and invoked fair-use principles; court outcomes and claims varied by case and jurisdiction.

  Source
• 2025Privacy concerns over using AI conversations for personalizationControversy

  Meta announced that interactions with its AI features would be used to personalize content and advertising recommendations. Privacy advocates criticized the policy’s implications for sensitive conversational data, cross-product profiling, transparency, and user control.

  Source
• 2024News-summary attribution and publisher concernsControversy

  Independent testing reported that Meta AI summarized news stories with limited attribution and, in some cases, reproduced wording or supplied incomplete links. The issue was especially contentious in Canada, where Meta had restricted news links following a dispute over publisher compensation and regulation.

  Source
• 2022Galactica generates unreliable and offensive scientific materialControversy

  Meta AI’s Galactica demonstration was withdrawn after users reported fabricated citations, inaccurate scientific claims, racist material, and other offensive outputs. The incident drew attention to the danger of presenting fluent model output in a scientific register without dependable verification.

  Source

• 2026Meta introduces Muse Spark for Meta AI

  Meta Superintelligence Labs introduced Muse Spark as the first model in a new Muse series and said it would power the Meta AI app and website before expanding to Meta’s messaging, social, and wearable products.

  Product launchProduct generation
  Source
• 2025Meta launches a standalone Meta AI application

  Meta launched a standalone Meta AI app built with Llama 4, adding personalized conversations, voice interaction, image tools, a Discover feed, and continuity with meta.ai and Meta’s AI glasses.

  Product launch
  Source
• 2025Meta reorganizes advanced AI work under Meta Superintelligence Labs

  Meta consolidated FAIR, language-model development, and other AI research and product teams under Meta Superintelligence Labs as part of a strategy focused on advanced and personal AI.

  Leadership change
  Source
• 2025Meta announces personalization based on AI interactions

  Meta announced that interactions with its generative AI features would become signals for personalizing content and advertising recommendations, with the change scheduled to take effect on December 16, 2025.

  Pricing
  Source
• 2024Meta AI expands internationally with Llama 3

  Meta expanded Meta AI beyond the United States across WhatsApp, Instagram, Facebook, Messenger, and the meta.ai website, while adding image generation, web search, and broader conversational assistance.

  Product launchCampaign
  Source
• 2024Meta AI adds vision features to Ray-Ban Meta glasses

  Meta rolled out Meta AI with Vision, allowing supported Ray-Ban Meta smart glasses to use camera input for tasks such as identifying objects, translating text, and answering questions about a wearer’s surroundings.

  Product launchProduct generation
  Source
• 2024Meta adds voice and photo understanding to its AI products

  Meta announced voice conversations, photo-based question answering, and expanded multimodal capabilities for Meta AI across Messenger, Facebook, WhatsApp, and Instagram.

  Product launch
  Source
• 2023Meta AI introduces the consumer assistant

  Meta introduced a conversational Meta AI assistant based on its Llama models and began integrating it into its consumer products and services.

  Product launch
  Source

1. Meta AI
2. Meta Superintelligence Labs
3. Meet Your New Assistant: Meta AI, Built With Llama 3
4. New Ray-Ban | Meta Smart Glasses Styles and Meta AI Updates
5. Meta’s AI Products Just Got Smarter and More Useful
6. Introducing the Meta AI App: A New Way to Access Your AI Assistant
7. Seamless Communication
8. Improving Your Recommendations on Our Apps With AI at Meta
9. Personal Superintelligence for Everyone
10. Introducing Muse Spark: Meta’s Most Powerful Model Yet
11. New Meta AI demo writes racist and inaccurate scientific literature, gets pulled
12. How Meta AI’s news summaries could damage the industry
13. Judge dismisses authors’ copyright lawsuit against Meta over AI training