Meta AI

Meta AI is the name Meta Platforms uses for two related but distinct things: the company's artificial intelligence research and engineering organization, and the consumer assistant it ships inside Facebook, Instagram, Messenger, WhatsApp, a standalone app, Meta Quest headsets, and its AI glasses. The organizational sense is the older one. Meta (then Facebook) opened its research lab in December 2013 under Yann LeCun and branded a decade of published work, open datasets, and open-weight models under the "Meta AI" and "AI at Meta" labels. The product sense dates from September 27, 2023, when Meta launched a general-purpose assistant with that name.^[2][13][33]

Both senses are covered here. The research side produced PyTorch, the LLaMA language models, the Segment Anything family, and a long line of speech, translation, and self-supervised vision systems. The product side reached what Meta said was a billion monthly users across its apps by May 2025, a figure no competing assistant has claimed.^[54] Since mid-2025 the research and product work has been consolidated inside Meta Superintelligence Labs (MSL), an organization created after Meta paid about $14.3 billion for a minority stake in Scale AI and recruited its co-founder Alexandr Wang as the company's first Chief AI Officer.^[16][38] Meta FAIR, the original laboratory, remains a unit inside that structure.^[13][37]

The technology under the assistant has changed repeatedly. The 2023 launch used a custom model built on Llama 2; later versions used Llama 3, Llama 4, and, from April 2026, the Muse family developed by MSL. Meta released downloadable weights for several Llama generations under its own community licenses, which the Open Source Initiative says do not meet the Open Source Definition. Muse models are served through Meta products and a paid API rather than distributed as weights.^[9][26][27]

Scope and terminology

Meta Platforms, Inc. is the legal registrant that reports Facebook, Instagram, Messenger, and WhatsApp as its "Family" of products. Meta AI is one of the services provided within that corporate group. The product name is not the registrant's legal name and should not be used as a synonym for Meta Platforms as a whole.^[1]

Four related uses of the name are worth distinguishing:

Scroll sideways for more →

The distinction matters because many systems associated with Meta predate the assistant. FAIR produced research in computer vision, translation, and representation learning for a decade before any chatbot shipped, and Meta decentralized several applied AI teams into product groups in 2022. Those histories explain the technology and personnel behind Meta AI, but they do not make every Meta machine learning project a feature of the assistant.^[13][14] An example on the applied side is GEM, Meta's Generative Ads Recommendation Model, the foundation model behind ads recommendations across Instagram and Facebook; in an August 3, 2026 engineering post, Meta said GEM now trains at LLM scale on several thousand of its latest-generation GPUs.^[102]

Organizational history

FAIR and the first decade, 2013 to 2022

Facebook announced its research laboratory on December 9, 2013, naming LeCun, then a professor at New York University, as its director. The lab began at three sites: Menlo Park, London, and New York City. A Paris laboratory followed in 2015, and Meta later added Montreal, Tel Aviv, Seattle, and Pittsburgh.^[13][33] The name changed over time from Facebook AI Research to Fundamental AI Research, keeping the FAIR acronym.^[13]

FAIR's charter emphasized long-horizon research, publication, and open release. In practice that meant papers, code, datasets, and model weights rather than shipped features. The lab's most consequential release was probably PyTorch, published in 2017 and moved to an independent PyTorch Foundation under the Linux Foundation in September 2022; it became one of the two dominant deep learning frameworks and is now used by most of the industry, including Meta's competitors.^[13][80] Other durable outputs include the Detectron2 object detection library, the FAISS similarity search library, RoBERTa, BART, and the wav2vec speech models.^[13]

In January 2018 Meta restructured its AI work. LeCun moved from director to Chief AI Scientist, a research-focused role, and Jerome Pesenti joined as Vice President of AI with oversight of both the research lab and the applied machine learning group.^[35] LeCun said publicly that he wanted to step back from running the laboratory day to day.^[35]

The second restructuring came on June 2, 2022, when CTO Andrew Bosworth published a post describing how Meta would distribute applied AI teams into the product organizations that used them, while moving FAIR into Reality Labs Research.^[14] Reporting through 2023 and 2024 described FAIR's internal profile as narrowing during this period, with the highest-profile Llama releases coming from a separate product group rather than from the laboratory.^[13]

The GenAI product organization, 2023 to 2025

On February 27, 2023, Zuckerberg announced a new top-level product group focused on generative AI, pulling together teams that had been scattered across the company. It reported to Chief Product Officer Chris Cox and was led by Ahmad Al-Dahle, a vice president who had joined Meta in 2020 after sixteen years at Apple.^[34] This group, generally called GenAI, built the Meta AI assistant and shipped the Llama 2, Llama 3, and Llama 4 model families. It was organizationally separate from FAIR, which is why describing FAIR as "the chatbot division" is inaccurate.^[13][34]

Two senior departures preceded the 2025 reorganization. Joelle Pineau, the vice president who ran FAIR, announced on April 1, 2025 that she would leave; her last day was May 30, 2025, and she joined Cohere as Chief AI Officer that August.^[36] On May 8, 2025, Meta named Rob Fergus, a computer vision researcher who had spent about five years as a research director at Google DeepMind and who had helped start FAIR alongside LeCun, to lead the laboratory.^[37]

The 2025 reorganization and Meta Superintelligence Labs

The largest change came in mid-2025. In June, Meta agreed to invest about $14.3 billion for a stake of roughly 49 percent in Scale AI, the data labeling company, valuing it above $29 billion. The structure was a non-controlling minority investment rather than an acquisition. Wang stepped down as Scale's chief executive to join Meta.^[16]

On June 30, 2025, Zuckerberg sent an internal memo creating Meta Superintelligence Labs and naming Wang as Meta's first Chief AI Officer, with former GitHub chief executive Nat Friedman leading AI products and applied research. The memo named roughly eleven external recruits drawn from OpenAI, Google DeepMind, Anthropic, and other laboratories.^[38] On July 25, 2025, Wang announced that Shengjia Zhao, a former OpenAI researcher who had worked on ChatGPT and on OpenAI's reasoning models, would be MSL's Chief Scientist.^[39] In prepared remarks for the second quarter of 2025, Mark Zuckerberg described MSL as combining the foundation model teams, the product teams, FAIR, and a new laboratory focused on the next generation of models.^[15]

Zuckerberg set out the strategy publicly in a short essay, "Personal Superintelligence," on July 30, 2025. He argued that Meta's aim differed from rivals who framed advanced AI mainly as a way to automate economically valuable work, and that Meta would deliver a personal superintelligence through devices such as glasses. The essay also struck a more cautious note on open release, suggesting Meta would be more selective about what it published as systems grew more capable.^[41]

The recruiting drive that staffed MSL became a story in its own right. OpenAI chief executive Sam Altman said publicly in June 2025 that Meta had offered signing bonuses as large as $100 million to some of his staff; subsequent reporting described multiyear packages for a handful of senior researchers running into the tens or hundreds of millions of dollars. Meta executives disputed how those figures were characterized. None of the individual terms have been disclosed by Meta, and the numbers in circulation come from interested parties and press accounts rather than filings.^[38] The contrast between those packages and the layoffs that followed within months drew persistent internal and external criticism.^[18]

An August 2025 report said Meta planned to divide MSL into four groups: a model-focused unit called TBD Lab, FAIR, a product group that included the Meta AI assistant, and an infrastructure group. Reuters said at the time that it could not independently verify the reported plan; the four-part structure has been widely repeated but Meta has not published an organization chart confirming it.^[17] On October 22, 2025, Meta confirmed cuts of about 600 positions in the AI organization, affecting FAIR, product AI, and infrastructure teams while sparing the newer frontier model group.^[18]

Yann LeCun confirmed on November 19, 2025 that he would leave Meta after about twelve years to start a company focused on world models, the research direction he had long argued was more promising than scaling language models. His venture is Advanced Machine Intelligence (AMI) Labs, co-founded with Alexandre LeBrun and headquartered in Paris.^[40] Meta has not named a company-wide successor to the Chief AI Scientist title; Zhao holds the Chief Scientist role at MSL and Fergus leads FAIR.^[37][39]

2026: Applied AI Engineering, layoffs, and the current shape

Meta made two further structural changes in 2026 that are frequently confused with MSL.

In January 2026, Zuckerberg announced Meta Compute, a top-level organization for planning, building, and running the company's data center capacity. It is co-led by head of global infrastructure Santosh Janardhan and Daniel Gross, the former Safe Superintelligence chief executive who had joined Meta in 2025.^[69] Meta Compute is an infrastructure function, not a model laboratory, and both Zuckerberg and CFO Susan Li referred to it by name on the 2026 earnings calls.^[56][57]

Around March 2026, Meta created an Applied AI Engineering organization under Maher Saba, a longtime Reality Labs vice president, reporting to CTO Andrew Bosworth rather than to Wang. Reuters reported in April 2026 that engineers were being transferred into the group and that the transfers were not optional.^[19] By June 2026 the unit had roughly 6,500 engineers and product managers, whose work centered on producing training data, coding problems, and evaluations for Meta's models; TechCrunch reported that employees described the assignment in strongly negative terms and that Zuckerberg acknowledged in an internal memo that recent changes had "caused distress."^[91] Applied AI Engineering sits outside MSL and should not be conflated with either FAIR or the frontier model group.

On May 20, 2026, Meta began notifying roughly 8,000 employees of layoffs, the largest company-wide reduction since the 2022 and 2023 efficiency campaigns. Meta booked $1.2 billion in severance expense in the second quarter and ended June 2026 with 75,472 employees, down 1 percent year over year.^[55][56]

The table below summarizes the leadership positions Meta or its executives have confirmed. Personnel reporting in this area has frequently been premature, so only confirmed roles are listed.

Scroll sideways for more →

The Llama family

Llama was Meta's open-weight language model program from 2023 through 2025. It made Meta the most significant corporate publisher of downloadable frontier-adjacent model weights and shaped a large part of the open ecosystem, including quantized runtimes, fine-tuning toolchains, and derivative models from other laboratories.

Llama 1 and the leak

Meta published the first LLaMA models on February 24, 2023 at 7B, 13B, 33B, and 65B parameters. The 65B and 33B models were trained on 1.4 trillion tokens and the 7B on one trillion. Meta released them under a non-commercial research license with access granted case by case to academic, government, civil society, and industry researchers who applied through a form.^[42] The gate did not hold. About a week after Meta began accepting applications, a 4chan user posted a BitTorrent link to the complete model package. Meta filed takedown requests against repositories hosting the weights on GitHub and Hugging Face, but redistribution continued, and within days developers were running the models on consumer hardware and building derivatives such as Stanford's Alpaca.^[100] The leak is usually credited with accelerating the open-weight ecosystem that Meta went on to embrace deliberately.

Llama 2 and the community license

Llama 2, released July 18, 2023, was the pivot to open weights with commercial use permitted. The family shipped at 7B, 13B, and 70B parameters with a 4,096-token context window, trained on two trillion tokens with a September 2022 pretraining cutoff. The 70B model used grouped-query attention for cheaper inference. A 34B model appeared in the accompanying paper but was never released.^[43] The license was the first "LLAMA 2 COMMUNITY LICENSE AGREEMENT," a bespoke agreement rather than a standard open-source license.^[43] Meta shipped Code Llama in August 2023 as a code-specialized derivative, and Purple Llama, a set of safety tools and classifiers, that December.

The Llama 3 generation

Llama 3 arrived April 18, 2024 at 8B and 70B with an 8,192-token context, trained on more than 15 trillion tokens.^[44] Llama 3.1 followed on July 23, 2024, adding a 405B dense model and extending context to 128,000 tokens across the family; it was the largest openly downloadable model of its time.^[45] Llama 3.2 on September 25, 2024 added small on-device models at 1B and 3B plus the first Llama vision models at 11B and 90B.^[46] Llama 3.3, released December 6, 2024, was a 70B instruction-tuned model that Meta positioned as matching the 405B model's quality on many tasks at far lower cost.^[47]

Llama 4 and its reception

Llama 4, released April 5, 2025, moved the family to a mixture of experts architecture and native multimodality. Llama 4 Scout and Maverick each activate 17 billion parameters per token: Scout has 16 experts and 109 billion total parameters with a claimed 10 million token context, Maverick has 128 experts and about 400 billion total parameters with a 1 million token context. Both were trained on more than 30 trillion tokens with an August 2024 data cutoff and support twelve languages.^[48][49] Meta previewed a third model, Llama 4 Behemoth, with 288 billion active parameters, 16 experts, and nearly two trillion total parameters, describing it as still training and used as a teacher model for distilling the smaller two.^[48] Behemoth was never released, and Meta has not published a formal cancellation.

The launch went badly in two ways. Independent testers reported that the deployed Scout and Maverick models did not behave like the version Meta had submitted to the LMArena leaderboard, which was an unreleased variant labeled "Llama-4-Maverick-03-26-Experimental." Meta's VP of generative AI, Ahmad Al-Dahle, denied on April 7, 2025 that Meta had trained on test sets, writing "that's simply not true and we would never do that," and attributed variable quality to implementations needing to stabilize.^[92] In January 2026, after announcing his departure, LeCun told the Financial Times that Meta researchers had "fudged a little bit" by reporting results from different Llama 4 variants on different benchmarks and presenting them together.^[93]

Release table

Scroll sideways for more →

Meta has not released a new open-weight Llama generation since April 2025. As of August 2026 the meta-llama organization on Hugging Face lists nothing newer than the Llama 4 models, and llama.com redirects to Meta's developer site rather than to a model catalogue of its own.^[49] The Llama name still appears in Meta's product line, notably the Llama API and Llama Stack, but the frontier work moved to Muse.^[9][53]

Adoption

Meta reported 650 million cumulative Llama downloads in December 2024, one billion in March 2025, and 1.2 billion at its first LlamaCon developer conference on April 29, 2025.^[51][52] Download counts are a coarse metric: they count repository pulls across many mirrors and quantizations rather than distinct deployments. At LlamaCon, Meta also previewed the Llama API, a hosted inference service, and promoted Llama Stack, its reference application framework.^[53]

Licensing and the open-weight debate

Why Meta published weights

Meta set out its reasoning most fully in a letter Zuckerberg published on July 23, 2024, the day of the Llama 3.1 release, titled "Open Source AI Is the Path Forward." The argument had three strands. The commercial one was that Meta does not sell model access as its primary business, so unlike a laboratory whose revenue depends on API calls, it loses little by giving weights away and gains from an ecosystem that optimizes, ports, and debugs its models for free. The competitive one was a comparison to Linux and to Meta's own experience with closed mobile platforms: Zuckerberg argued that a company whose products run on someone else's proprietary stack is permanently constrained, and that an open standard prevents any one vendor from holding that position in AI. The safety one was that open models are easier to scrutinize, and that concentrating capability in a few closed laboratories is itself a risk.^[77]

That position had practical effects beyond marketing. Llama weights became the default starting point for fine-tuning research, for quantized local inference on consumer hardware, and for national and regional model efforts that could not afford to pretrain from scratch. It also gave Meta a policy argument in Washington and Brussels, where the company consistently opposed rules that would restrict the distribution of model weights. Meta's later retreat from open frontier releases, described below, weakened that position without formally abandoning it.^[41][56]

What the community licenses say

Every Llama generation from Llama 2 onward shipped under a Meta-authored "Community License" rather than a standard open-source license such as Apache 2.0 or MIT. The core terms are consistent across generations. The Llama 4 Community License grants a non-exclusive, worldwide, royalty-free right to use, modify, and distribute the model materials, subject to conditions:^[26][49]

• A licensee whose products had more than 700 million monthly active users in the calendar month before the version's release date must request a separate license from Meta, which Meta may grant at its discretion.
• Distributions must display "Built with Llama" and include the attribution notice, and derivative model names must begin with "Llama."
• Use must comply with the Llama Acceptable Use Policy, which prohibits enumerated categories of use.
• Meta defines the model materials as its proprietary material rather than placing them in the public domain.

The Open Source Initiative position

The Open Source Initiative has repeatedly said that these terms do not satisfy the Open Source Definition. In a February 18, 2025 post by Jordan Maris, OSI listed three failures: the license fails freedom 0 (the freedom to use the model for any purpose), fails point 5 of the Open Source Definition by discriminating against users, and fails point 6 by restricting fields of endeavor. The post also noted that the Llama 3.x licenses added restrictions beyond those in Llama 2.^[27] "Open weight" is the less ambiguous description of what Meta actually published, and is the term the wiki uses for the concept at open weights; see also open-source AI for the broader dispute.

Meta's own usage has been inconsistent. Zuckerberg has described Llama as open source in public remarks and on earnings calls, while Meta's license text and model cards use "community license" and "Llama Materials."^[26][56]

EU restrictions and regulatory friction

The Llama 3.2 Community License did not grant rights under section 1(a) to individuals domiciled in, or companies with a principal place of business in, the European Union with respect to the multimodal models in that release. Text-only Llama 3.2 models were unaffected, and the restriction did not extend to end users of products built on the restricted models.^[50] The restriction followed a period of uncertainty over how Meta could train on European user data, discussed below.

Meta also declined to sign the European Commission's voluntary Code of Practice for general-purpose AI models. Chief global affairs officer Joel Kaplan announced the refusal on July 18, 2025, writing that the code "introduces a number of legal uncertainties for model developers, as well as measures which go far beyond the scope of the AI Act." Meta was the first major developer to say publicly that it would not sign; OpenAI and Mistral did sign.^[83] For the underlying regulation, see EU AI Act.

The 2025 and 2026 shift

Meta's open-weight posture narrowed sharply after the Llama 4 launch. The "Personal Superintelligence" essay in July 2025 said Meta would be more careful about what it opened.^[41] MSL's first models were not released as weights at all. Asked directly on the July 29, 2026 earnings call about returning to open source, Zuckerberg said Meta had "always basically said that we were going to do a mix of open and closed," that ramping up MSL made open releases more work because an open model has to be "more well-rounded" rather than tuned for internal use, and that Meta expected to "get back to releasing some open source models at some point soon."^[56] He also rejected the idea that Meta could rely on other companies' open-weight models instead of building its own: "right now the open source models are not as strong as the frontier models."^[56]

Meta did continue publishing research weights under non-commercial terms during this period. The Code World Model, a 32-billion-parameter research model released by FAIR in September 2025, shipped under a FAIR Non-Commercial Research License.^[76]

The Muse family

Muse is the model family developed by Meta Superintelligence Labs and the first generation of Meta frontier models distributed as a hosted service rather than as weights.

Muse Spark, announced April 8, 2026, is a natively multimodal reasoning model with tool use, visual chain of thought, and multi-agent orchestration. Meta said MSL had "rebuilt our AI stack from the ground up" over the preceding nine months and that Muse Spark reached comparable capability to Llama 4 Maverick with "over an order of magnitude less compute." A "Contemplating" mode running parallel agents scored 58 percent on Humanity's Last Exam and 38 percent on FrontierScience Research by Meta's own measurement. The model shipped through meta.ai and the Meta AI app with a private API preview, and the app gained an Instant mode and a Thinking mode.^[9] A separate technical report by Meta researchers documented predeployment safety evaluations under Meta's Advanced AI Scaling Framework.^[10]

Muse Spark 1.1 followed on July 9, 2026, with a 1 million token context, gains in tool and computer use, coding, and multimodal understanding, and the public preview of the Meta Model API, described as OpenAI-compatible.^[11] Reporting at the time described the API as Meta's first direct paid developer access to a Muse model.^[28] On July 24, 2026, Meta began rolling out agent-like functions in selected markets, including recurring briefings, research and slide creation, and optional connections to email and calendar services, alongside an incognito chat mode.^[12]

Muse Image launched July 7, 2026 as MSL's first image generation model, replacing third-party image models Meta had licensed for the assistant. Meta previewed a companion video model, Muse Video, the same day, describing it as sharing the image model's pretraining base and adding native audio, with availability "coming soon."^[29] The Muse Image launch included a feature that let a user mention a public Instagram account so its public images could be used as references; Meta withdrew that feature on July 10, 2026 after privacy and consent criticism, leaving the model available for other tasks.^[30]

On the July 2026 earnings call Zuckerberg said Meta had seen "a 60% increase in the number of people interacting with the assistant each day" since rebuilding Meta AI around Muse Spark, and that Meta had begun using Muse models for content understanding tasks such as video topic classification.^[56]

Scroll sideways for more →

Research and model output beyond Llama

Meta's research organizations have produced a large body of work that never became a consumer product. The list below is representative rather than exhaustive; individual pages carry the technical detail.

Vision and representation learning

FAIR's self-supervised vision line runs from DINO (2021), which showed that self-supervised vision transformers learn segmentation-like attention without labels, through DINOv2 (2023) and DINOv3 (2025), each scaling the recipe and stabilizing dense features over longer training schedules.^[74] I-JEPA (2023) applied LeCun's joint embedding predictive architecture to images, predicting representations rather than pixels.

The Segment Anything Model (SAM), released in April 2023 with the SA-1B dataset of about 1.1 billion masks over 11 million images, made promptable segmentation a general-purpose primitive.^[70] SAM 2 extended it to video in July 2024.^[71] SAM 3, released November 19, 2025, added promptable concept segmentation: given a noun phrase such as "yellow school bus" or an image exemplar, it returns masks and identities for every matching instance at once, where earlier versions predicted one object per prompt. Meta released it alongside SAM 3D for single-image reconstruction and a public Segment Anything Playground.^[72] SAM 3.1, published March 27, 2026, is a drop-in replacement that tracks up to 16 objects in a single forward pass and roughly doubles throughput.^[73]

Meta's answer to OpenAI's CLIP was MetaCLIP (2023), which reverse-engineered and published a reproducible data curation recipe rather than treating the training set as a trade secret; MetaCLIP 2 (2025) extended the recipe to worldwide, non-English web pairs and reported that multilingual training improved English accuracy rather than degrading it.^[98] Hiera (ICML 2023) argued that much of the added complexity in hierarchical vision transformers was unnecessary once strong pretraining was applied, and removed it. Perception Encoder and the Perception Language Model, published in April 2025, packaged Meta's image and video encoders and a family of open vision-language models built on Llama 3 backbones.^[97] Sapiens (2024) covered human-centric vision tasks.

Speech, audio, and translation

wav2vec (2019) and wav2vec 2.0 (2020) established self-supervised pretraining for speech recognition and became the standard starting point for low-resource languages. No Language Left Behind (2022) delivered translation across 200 languages. Massively Multilingual Speech (2023) extended speech recognition to more than 1,100 languages and language identification to more than 4,000. SeamlessM4T (2023) unified speech-to-speech, speech-to-text, text-to-speech, and text-to-text translation in one model. SpiRit-LM (2024) interleaved speech and text tokens in a single language model so that a system could switch modalities mid-sequence.^[13]

On the generative audio side, Voicebox (2023) handled text-guided speech infilling and editing, and AudioCraft bundled MusicGen, AudioGen, and the EnCodec neural codec. Audiobox (2023) generalized voice and sound generation with natural language prompts. Several of these releases shipped with watermarking or detection tooling, reflecting Meta's stated position that generative audio needs provenance signals.

World models and embodied AI

Meta's world model work is the research direction LeCun championed and then left to pursue independently. V-JEPA (February 2024) trained a video model to predict in representation space rather than pixel space; V-JEPA 2, released June 11, 2025, scaled the recipe to roughly 1.2 billion parameters on more than one million hours of video and was used for zero-shot robot planning in unfamiliar environments.^[75] Meta Motivo (December 2024) is a behavioral foundation model for controlling a simulated humanoid body across tasks it was not explicitly trained on.

AI Habitat is Meta's simulation platform for embodied AI, with Habitat 3.0 (2023) adding humanoid avatars and human-robot collaboration scenarios. Ego4D (2022) and Ego-Exo4D provided large egocentric video datasets recorded by camera-wearing participants, which underpin much of the perception work behind AI glasses. Project Aria is the research glasses program that supplies that data, distributed to external research partners rather than sold. Droidlet (2021) was an earlier open platform for building embodied agents.^[13]

Language, code, and architecture research

OPT (May 2022) was an early attempt at an openly released 175-billion-parameter model, notable as much for publishing its training logbook as for the weights themselves. Toolformer (2023) showed a model teaching itself when to call external APIs. MEGABYTE (2023) and the Byte Latent Transformer (2024) both attacked tokenization, the latter replacing tokens with dynamically sized byte patches. The Large Concept Model (2024) moved prediction from the token level to a sentence-level embedding space. Self-Taught Evaluator (2024) trained a model to judge outputs using synthetic preference data instead of human annotation. LLM Compiler (2024) adapted Code Llama to compiler optimization and disassembly. LIMA (2023) argued that a small, carefully curated instruction set could substitute for large-scale RLHF. HalluLens (2025) is Meta's benchmark taxonomy for hallucination. Atlas was a retrieval-augmented few-shot language model.^[13]

The Code World Model, released by FAIR in September 2025, is a 32-billion-parameter model trained on Python interpreter execution traces and agentic Docker interactions so that it can reason about what code does when it runs rather than only about how it looks. It shipped under a non-commercial research license.^[76]

Generative media research produced Make-A-Scene and Make-A-Video (2022), CM3leon (2023), the Emu family including Emu Video and Emu Edit, Chameleon (2024), an early-fusion mixed-modal architecture, and Movie Gen (2024). ImageBind (2023) bound six modalities into a single embedding space using only image-paired data.

Chatbots before the assistant, and what went wrong

Meta shipped conversational research systems for years before Meta AI. BlenderBot 1 (2020) and 2 (2021) were open-domain dialogue agents. BlenderBot 3, released as a public demo on August 5, 2022, was designed to search the web and learn from live conversations, and within days it was reported producing antisemitic tropes, false claims about the 2020 United States election, and criticism of Zuckerberg. Meta required users to acknowledge that the bot was for research and entertainment and could make untrue or offensive statements, and it left the demo online to collect feedback.^[101] The episode is a standard example of why open-ended deployment of a research chatbot is difficult.

Galactica, a model trained on scientific literature including tens of millions of papers, launched as a public demo on November 15, 2022 and was withdrawn three days later. Meta had promoted its ability to summarize literature, solve mathematical problems, generate encyclopedia articles, write scientific code, and annotate molecules and proteins. Researchers instead demonstrated that it produced fluent, confidently formatted, and factually wrong scientific text, complete with plausible-looking citations to work that did not exist, on subjects the model should have refused outright.^[86] The withdrawal came two weeks before ChatGPT's launch and is routinely cited in retrospective accounts of why Meta was slower than its rivals to ship a consumer assistant.

Meta's game-playing research includes Pluribus (2019), which beat professional players at six-player no-limit poker, and CICERO (2022), which combined a language model with strategic reasoning to play the negotiation game Diplomacy at human level. On the science side, ESMFold and the ESM Metagenomic Atlas (2022) predicted protein structures directly from sequence, and the Open Catalyst Project applied machine learning to catalyst discovery.^[13]

Products

The Meta AI assistant

Meta announced the assistant on September 27, 2023, releasing it in beta in the United States through WhatsApp, Messenger, and Instagram. Users could start a direct conversation or invoke it in a group chat. The launch version used a custom conversational model built on Llama 2 technology, retrieved current information through a search partnership with Bing, and generated images with Meta's Emu system. Support for Ray-Ban Meta smart glasses and Meta Quest was announced alongside the chat interfaces.^[2]

Meta did not present the launch model as reliable. Its accompanying safety statement said generative systems could return inaccurate or inappropriate output and described the release as a staged beta, listing red-team testing, output filtering, and model-specific safeguards as mitigations.^[3]

In April 2024, Meta replaced the underlying model with Llama 3 and expanded access to Facebook, Instagram, Messenger, WhatsApp, and a web interface.^[4] Voice conversations and image input followed in September 2024 using models from the Llama 3.2 release.^[5] The European rollout began in March 2025 across 41 countries and 21 overseas territories, starting with text chat in six languages; Meta said the delay reflected regulatory work and that advanced functions would arrive separately.^[6]

Meta released the first standalone Meta AI mobile app on April 29, 2025, built with Llama 4, incorporating the device-management functions of the former Meta View app and synchronizing conversations with meta.ai. It introduced a Discover feed for prompts and generated material, which Meta said required an explicit publish action.^[7] The April 2026 Muse Spark release rebuilt the app around Instant and Thinking modes.^[9]

Availability has never been uniform. Meta's own user guide describes three access routes (the app or website, direct chats inside the four messaging surfaces, and voice on supported Quest devices and AI glasses) and warns that supported file types and features vary by channel and country.^[8]

The lineage can be summarized without treating each product update as a new assistant:

Scroll sideways for more →

Model names do not fully specify the deployed system. Search providers, retrieval indexes, image and speech models, safety classifiers, system instructions, personalization logic, and product tools also determine what a user receives, and Meta can change those components without renaming the assistant.

Personalization, memory, and agent features

Meta designed the standalone app around personalization. Users can tell the assistant to remember preferences, and the system can preserve details from earlier conversations. In supported markets it can also draw on profile information and content a person has chosen to share on connected Meta accounts, with Facebook and Instagram linked through Accounts Center.^[7] Meta's privacy documentation says a chat may store the user's messages, the assistant's replies, and details saved from the conversation, and provides commands and account tools to inspect, download, reset, or delete some of that information. Deleting a visible chat in one app is not always the same operation as deleting Meta's copy of the conversation.^[21]

Muse Spark introduced a more tool-oriented system. Meta says the assistant can decompose tasks, call tools, and coordinate multiple model instances; Muse Spark 1.1 added longer-running workflows and optional connections to outside services such as email and calendar.^[9][11][12] These functions place Meta AI within the category of AI agents, but the label does not imply autonomy without limits: a tool can fail, an outside service can return incomplete data, and a model can select an incorrect action.

AI Studio, characters, and Vibes

Meta AI is not the only conversational product in the family. Meta AI Studio, rolled out to United States users on July 29, 2024, lets anyone build customizable AI characters without writing code and deploy them in Instagram, Messenger, WhatsApp, or on the web. Creators can also build an AI version of themselves that answers common messages.^[88] These character AIs are governed by different policies from the general assistant and were the focus of most of the 2025 and 2026 safety scrutiny described below.

Vibes, launched September 25, 2025, is an AI video feed inside the Meta AI app and on meta.ai where users generate and remix short videos from text prompts and cross-post them to Instagram Reels or Facebook Stories.^[87]

In 2026 Meta added Meta One, a paid subscription bundling AI features across its apps, and rolled out Meta Business Agents globally on WhatsApp and Messenger, with more than one million businesses using them weekly by the second quarter.^[56]

Smart glasses

Glasses are the hardware Meta associates most closely with its assistant. Ray-Ban Meta glasses, built with EssilorLuxottica, launched in September 2023 with a camera, open-ear speakers, a microphone array, and Meta AI voice access. Meta added an Oakley line in 2025 and shipped Meta Ray-Ban Display, its first heads-up display product, on September 30, 2025 at $799 including a Neural Band wristband that reads muscle signals at the wrist.^[89]

EssilorLuxottica said in February 2026 that it had sold more than seven million smart glasses in 2025, more than tripling 2024 volumes.^[90] On the first-quarter 2026 earnings call Zuckerberg said the number of people using the glasses daily had tripled year over year, and on the second-quarter call he said Meta had released its own line of Meta Glasses with EssilorLuxottica, the first to ship with Muse Spark out of the box.^[56][57] Reality Labs revenue was $431 million in the second quarter of 2026, up 16 percent, which Meta attributed to AI glasses growth partly offset by lower Quest headset sales.^[55]

Scale, monetization, and how the numbers should be read

Meta AI's headline user figures are the largest of any assistant, and they are also the hardest to compare with rivals. Zuckerberg said at Meta's annual shareholder meeting on May 28, 2025 that the assistant had one billion monthly active users, up from about 500 million in September 2024.^[54] That count includes anyone who interacts with Meta AI anywhere in the Family of Apps, including people who tap the assistant inside a Facebook search box or an Instagram direct message. It is not equivalent to a figure for a dedicated assistant app, and Meta has not published a comparable daily active number for the assistant alone. The relative measures Meta has disclosed since are more informative than the absolute ones: double-digit percentage increases in Meta AI sessions per user after the Muse Spark rollout in April 2026, and a 60 percent increase in daily interactions by July 2026.^[56][57]

Monetization arrived late and in several forms at once. Meta began using AI interactions as ad and content recommendation signals in December 2025, launched the Meta One subscription in 2026, opened the Meta Model API to paying developers in July 2026, and rolled out business agents priced through a mix of subscriptions and volume-based pricing. Zuckerberg has said he expects more of these products to move toward outcome-based pricing resembling Meta's ad auction, where businesses pay when the system produces a result, and that Meta is weighing selling raw compute alongside model access.^[24][56] Family of Apps "other" revenue, which includes WhatsApp paid messaging and subscriptions, passed $1 billion in a quarter for the first time in the second quarter of 2026.^[55]

The competitive position is genuinely mixed. Meta has by far the widest distribution and the deepest capital commitment, but its models arrived after the frontier had already been set by OpenAI, Google DeepMind, and Anthropic, and Zuckerberg described Muse Spark 1.1 on the July 2026 call as strong for its scale while acknowledging that Meta was still climbing toward more advanced models.^[56] Meta's own framing of the difference is strategic rather than technical: it is, in Zuckerberg's words, "the only major company building AI with the primary goal of putting superintelligence directly into people's hands," distributing capability through consumer apps and glasses rather than selling it centrally.^[56]

Infrastructure

Research SuperCluster

Meta announced the Research SuperCluster on January 24, 2022 as a machine for training large models on real production data. Phase one was 760 NVIDIA DGX A100 systems containing 6,080 A100 GPUs, connected by InfiniBand with 175 petabytes of Pure Storage FlashArray, 46 petabytes of cache, and 10 petabytes of FlashBlade. Meta expanded it to 16,000 A100 GPUs during 2022.^[64] RSC trained the first LLaMA models.^[42]

Open hardware

Meta has contributed its AI server and rack designs to the Open Compute Project, the open hardware effort it helped found in 2011. Grand Teton, announced at the OCP Global Summit on October 18, 2022, consolidated the CPU head node, GPU system, and switching that its predecessor Zion-EX had cabled together into a single chassis, roughly quadrupling host-to-GPU bandwidth and doubling network bandwidth and power envelope. The first version carried eight H100 GPUs; a 2024 revision added support for the AMD Instinct MI300X.^[65]

Catalina, unveiled at the October 2024 OCP Summit, is a two-rack, liquid-cooled pod built around NVIDIA's GB200 NVL72 platform, forming a single 72-GPU scaling domain. It is the first high-power implementation of Meta's Open Rack v3 standard, raising the per-rack ceiling to roughly 140 kW with 480-volt input converted to a 48-volt DC busbar.^[66]

MTIA

MTIA (Meta Training and Inference Accelerator) is Meta's family of custom ASICs, designed for its own data centers and not sold. Meta announced the first generation on May 18, 2023 as an inference part for the deep learning recommendation models behind ranking and ads, built on a TSMC 7nm process with an 8x8 grid of RISC-V based processing elements and a 25 W thermal design power. Meta's published results were modest: roughly 0.9x performance per watt against a GPU baseline across recommendation models, with individual operators reaching about 2x.^[60] The second generation, announced April 10, 2024, moved to 5nm, raised the clock from 800 MHz to 1.35 GHz, and reported 3.5x higher dense and 7x higher sparse compute than the first chip.^[61]

On March 11, 2026, Meta published a roadmap describing four chips shipping on a roughly six-month cadence, renaming the earlier parts MTIA 100 and MTIA 200 and building the new accelerators from reusable chiplets. Meta said it deploys hundreds of thousands of MTIA chips in production and develops them with Broadcom.^[62][63]

Scroll sideways for more →

Meta says HBM bandwidth rises about 4.5x and compute throughput about 25x across the span from MTIA 300 to MTIA 500, the latter measured from MTIA 300's MX8 rate to MTIA 500's MX4 rate.^[62] These are relative comparisons against Meta's own previous parts, published by Meta; no independent benchmarks of MTIA 300 or later were public as of mid-2026. The program also has a history of course changes: Meta cancelled a planned 2022 custom chip rollout and bought GPUs instead.^[60][62] The software stack is built on PyTorch, Triton, MLIR, and LLVM, with an MTIA plugin for vLLM, so models can move between MTIA and merchant GPUs in the same fleet.^[62]

GPU fleet, data centers, and capital expenditure

Meta continues to buy merchant silicon at very large scale alongside MTIA. In January 2024 Zuckerberg said Meta would have about 350,000 H100 GPUs by the end of that year and compute equivalent to roughly 600,000 H100s counting other accelerators.^[67] Reporting in February 2026 described an expanded multiyear NVIDIA agreement covering Blackwell and Rubin generation parts plus standalone Grace and Vera CPUs, followed days later by a roughly 6-gigawatt agreement with AMD.^[63]

The fleet is not only accelerators. On April 24, 2026, Meta signed a multibillion-dollar, multiyear agreement to deploy tens of millions of Amazon Web Services Graviton5 Arm CPU cores for the CPU-heavy orchestration layer around agent workloads: branching control flow, tool invocation, sandboxed code execution, and coordination across concurrent sub-agents. The deal is unusual because Meta operates its own data centers and designs its own silicon yet chose to source general-purpose compute from a rival hyperscaler; Meta framed it as diversification rather than a move to public cloud.^[99] Meta's silicon strategy also expanded by acquisition: it bought Rivos, a startup building RISC-V server CPUs paired with a data-parallel accelerator, in the autumn of 2025.^[96]

Meta disclosed two named multi-gigawatt clusters in July 2025: Prometheus, a roughly 1-gigawatt facility at New Albany, Ohio expected online in 2026, and Hyperion in Richland Parish, Louisiana, which Meta says will scale toward about 5 gigawatts over several years on a roughly 2,250-acre site.^[15] In January 2026 Zuckerberg created Meta Compute as a top-level organization for this buildout, saying Meta planned "tens of gigawatts this decade, and hundreds of gigawatts or more over time."^[69]

Meta increasingly finances that capacity off its own balance sheet. On July 28, 2026 it announced a joint venture with BlackRock for a $14 billion, 1-gigawatt campus in El Paso, Texas: BlackRock-managed funds hold 80 percent after contributing about $4.9 billion in cash, Meta holds 20 percent after contributing land and in-progress construction worth about $2.3 billion, and capacity is expected online in 2028.^[68] A comparable structure with Blue Owl Capital covers the Hyperion campus.^[68]

Capital expenditure figures should be attributed to Meta's own filings and earnings materials, since guidance has moved repeatedly within single years.

Scroll sideways for more →

Meta reported second-quarter 2026 revenue of $60.80 billion, up 28 percent, and net income of $15.85 billion, with total 2026 expenses guided to $165 billion to $169 billion. The quarter included $2.4 billion in legal charges and $1.2 billion in severance.^[55] CFO Susan Li said Meta expects to remain supply constrained on compute for the foreseeable future.^[56] Zuckerberg has described selling compute and model access to outside customers as a potential business line, saying Meta had received offers for compute at "a meaningful premium over what we paid for" it.^[56]

Privacy and data use

Training data and conversations

Meta says its generative models use a mixture of publicly available information, licensed data, and information from Meta products and services. This is broader than its separate statement about the original Llama 2 foundation models, which Meta said did not include Meta user data. A statement about one foundation model does not describe every consumer model, post-training dataset, retrieval source, or product interaction.^[20]

Meta's Privacy Center says interactions with its AI features may be used to develop and improve generative models. In the European region, Meta says it uses public posts and comments from adult accounts and interactions with AI under a legitimate-interests basis, while providing a process to object to future use.^[22]

The European path was contested. Meta notified Ireland's Data Protection Commission in March 2024 of plans to train on public content from adult EU and EEA accounts, then paused those plans in June 2024 after regulatory and advocacy pressure. Following a December 2024 European Data Protection Board opinion setting criteria for GDPR compliance in model training, Meta revised its approach with updated transparency notices, an improved objection mechanism, longer notification periods, de-identification, and output filtering. In April 2025 Meta said it would resume training on public posts and comments from adult EU users and on interactions with Meta AI, and that it would not use private messages between people for that purpose.^[23] The Llama 3.2 multimodal EU licensing restriction described earlier belongs to the same period of uncertainty.^[50]

Advertising and recommendations

On October 1, 2025, Meta announced that from December 16, 2025 it would use interactions with its generative AI features as signals for content and advertising recommendations in most regions. Meta said conversations touching religious views, sexual orientation, political views, health, racial or ethnic origin, philosophical beliefs, or trade union membership would not be used to show ads. Users can adjust ad and feed controls but cannot fully opt out of the change.^[24] Using interactions to target recommendations is a different operation from using them to improve a model, even though the same conversation may be relevant to both policies.

Users should not assume that a chat with an AI has the privacy properties of an end-to-end encrypted conversation with another person. The assistant must process prompts to generate a response, and connected tools can transmit data to additional services. The governing Meta AI terms differ among the United States and other regions, the European region, the United Kingdom, and Brazil.

The Discover feed and incognito chats

The 2025 standalone app's Discover feed produced a specific privacy failure. Soon after launch, reporters found personal and sensitive conversations that users had published to the public feed. Meta said chats were private by default and required an active share or publish action, but the volume and nature of the posts raised questions about whether users understood the interface and the audience.^[7][25] The episode did not show that chats were automatically public; it showed a usability risk at the point where a private interaction could become a public post.

Meta launched incognito chats on WhatsApp and in the Meta AI app during the second quarter of 2026, describing them as conversations "that even Meta can't see."^[56] That description applies to the surfaces and terms in force at the time and should not be generalized to every interface.

Reliability and safety

Meta AI generates responses rather than retrieving a guaranteed record from a database. It can produce a hallucination, misattribute a source, follow a misleading premise, or answer with unjustified confidence. Web access and citations help a user audit an answer but do not remove these risks, and Meta's product notices advise against relying on AI responses for important decisions.^[3][21]

Meta uses model training, system instructions, classifiers, filters, red-team exercises, and policy enforcement to reduce harmful output. These measures can lower measured risk without proving that a system is safe in every language, context, or adversarial interaction. Safety claims should identify who ran the evaluation, what model and deployment were tested, and which threat categories were included.

Meta's published governance document for catastrophic risk began as the Frontier AI Framework, released February 3, 2025. It takes an outcomes-led approach: Meta defines catastrophic outcomes and the threat scenarios that could produce them, then sets thresholds based on how much unique uplift a model provides toward those scenarios. A model assessed at the critical threshold is not released openly, and if the risk cannot be mitigated, Meta says it stops development.^[78] On April 7, 2026 Meta renamed the document the Advanced AI Scaling Framework and published version 2, which covers chemical and biological, cybersecurity, and loss-of-control risks, commits to at least annual review, and adds a category of outcomes that Meta considers potentially catastrophic but not yet rigorously measurable.^[79] The framework applies to frontier models, not to ordinary product behavior such as the chatbot content standards discussed below.

The Muse Spark Safety and Preparedness Report is a primary technical source written by Meta researchers. It evaluates chemical and biological, cybersecurity, and loss-of-control risks under Meta's Advanced AI Scaling Framework. The authors reported that pre-mitigation chemical and biological capabilities were likely in their high risk category, then concluded that safeguards reduced residual deployment risk to an acceptable level under the framework. That conclusion is Meta's own assessment, not an independent certification, and the report does not cover ordinary product harms.^[10]

Meta's 2025 annual filing separately lists possible harms from AI technologies, including illegal or inaccurate content, defamation, bias or discrimination, cybersecurity attacks, privacy violations, intellectual property disputes, and threats to safety or well-being. A corporate risk disclosure does not establish that any listed harm occurred in Meta AI, but it contradicts any claim that the company regards these risks as eliminated.^[1]

Policy and controversies

The content risk standards reporting and minors

In August 2025, Reuters published an examination of an internal Meta document titled "GenAI: Content Risk Standards," roughly 200 pages, which had been approved by Meta's legal and public policy teams and its chief ethicist. The standards had permitted chatbots to engage children in romantic or sensual conversations and to generate certain false or discriminatory material. The document covered Meta AI as well as character chatbots across Meta's apps. Meta confirmed its authenticity, said the child-directed examples were erroneous and inconsistent with its policies, and removed the relevant portions after Reuters asked about them.^[31] The report examined internal standards; it did not establish how often any covered behavior reached users.

The Federal Trade Commission subsequently included Meta Platforms and Instagram among seven recipients of compulsory information requests in a study of companion chatbots, seeking details about safety testing, youth protections, advertising, enforcement, and the handling of conversation data. The inquiry was a fact-finding study under the agency's 6(b) authority, not a finding that Meta had violated the law.^[32]

Meta announced parental controls for teen AI interactions in October 2025, including the ability to turn off one-on-one chats with AI characters, block specific characters, and see the topics teens discuss, with rollout beginning in early 2026 in the United States, United Kingdom, Canada, and Australia.^[85] On January 23, 2026, Meta went further and paused teen access to AI characters entirely, worldwide, including for accounts its age-prediction systems flagged as likely teens, saying access would resume when an updated experience was ready. Teens retained access to the general Meta AI assistant. The pause came shortly before Meta was scheduled to stand trial in Los Angeles alongside TikTok and YouTube over harms to children.^[84] Meta told investors in July 2026 that it continues to see scrutiny on youth-related issues in several markets and has youth-related trials scheduled for 2026 that "may ultimately result in a material loss."^[55]

Copyright litigation

Kadrey v. Meta Platforms is the most consequential AI copyright case against the company. Thirteen authors, including Richard Kadrey, alleged that Meta infringed their copyrights by training Llama on their books. On June 25, 2025, Judge Vince Chhabria of the Northern District of California granted partial summary judgment for Meta on the training claim, holding that training on the works was fair use on this record. The ruling was pointedly narrow: Chhabria wrote that it did not stand for the proposition that Meta's use of copyrighted material was lawful, but that these plaintiffs had failed to develop evidence for a market dilution theory that might have succeeded.^[81] A separate set of claims about Meta's use of BitTorrent to obtain the books, and specifically whether Meta re-uploaded or "seeded" copyrighted files while downloading them, was not resolved by that ruling and remained live.^[81] Details are covered at Kadrey v. Meta.

In July 2025, Strike 3 Holdings and Counterlife Media sued Meta in California federal court, alleging that Meta had used BitTorrent to download 2,396 of their copyrighted adult films between 2018 and 2025 and that the downloads were connected to AI training. Meta argued that the downloading was personal use by individuals, not corporate AI data collection. In June 2026 a judge allowed direct, vicarious, and contributory infringement claims based on the torrenting to proceed.^[82]

Research integrity episodes

Three episodes are commonly cited in discussions of Meta's research claims. Galactica's withdrawal in November 2022 showed the gap between fluent scientific formatting and factual accuracy. BlenderBot 3's public demo in August 2022 showed the difficulty of deploying a self-learning dialogue agent. The Llama 4 benchmark dispute in April 2025 involved a model submitted to a public leaderboard that was not the model users could download, followed in January 2026 by LeCun's statement to the Financial Times that Meta researchers had reported results from different variants together.^[92][93] Meta denied training on test sets and has not retracted the Llama 4 model cards.^[92]

Acquisitions and talent

Meta's 2025 and 2026 dealmaking has drawn scrutiny of its own. Beyond the Scale AI investment, Meta acquired or acqui-hired several AI companies, including the voice startups PlayAI and WaveForms, the RISC-V chip company Rivos in the autumn of 2025, and the Singapore-based AI agent company Manus, announced December 30, 2025 and reported at roughly $2 billion to $3 billion.^[94][96] In March 2026, Meta hired the team behind Dreamer, an agent-building startup led by David Singleton, in a talent deal under which Dreamer remained a separate entity and Meta took a non-exclusive technology license.^[95] Reported compensation packages for individual researchers in 2025, including claims of nine-figure offers, came from interested parties and press reporting rather than from Meta disclosures, and the terms of individual deals have not been published.^[38]

See also

• Meta Platforms
• Meta FAIR
• Meta Superintelligence Labs
• LLaMA
• Muse Spark
• MTIA
• Meta Compute
• Meta Ray-Ban AI Glasses
• Open weights

References

1. ^Meta Platforms, Inc., "Annual Report on Form 10-K for the year ended December 31, 2025," filed January 28, 2026. sec.gov/...meta-20251231
2. ^Meta, "Introducing New AI Experiences Across Our Family of Apps and Devices," September 27, 2023. about.fb.com/...ants-characters-and-creative-tools
3. ^Meta, "Building Generative AI Features Responsibly," September 27, 2023. about.fb.com/...generative-ai-features-responsibly
4. ^Meta, "Meet Your New Assistant: Meta AI, Built With Llama 3," April 18, 2024. about.fb.com/...meta-ai-assistant-built-with-llama-3
5. ^Meta, "Meta's AI Products Just Got Smarter and More Useful," September 25, 2024. about.fb.com/...metas-ai-product-news-connect
6. ^Meta, "Europe, Meet Your Newest Assistant: Meta AI," March 19, 2025. about.fb.com/...meet-your-newest-assistant-meta-ai
7. ^Meta, "Introducing the Meta AI App: A New Way to Access Your AI Assistant," April 29, 2025, updated September 12, 2025. about.fb.com/...ai-app-new-way-access-ai-assistant
8. ^Meta, "How to use Meta AI: Beginner's guide and best tips," June 5, 2026. ai.meta.com/...how-to-use-meta-ai
9. ^Meta Superintelligence Labs, "Introducing Muse Spark: Scaling Towards Personal Superintelligence," April 8, 2026. ai.meta.com/...introducing-muse-spark-msl
10. ^Cristina Menghini et al., "Muse Spark Safety & Preparedness Report," arXiv:2606.12429, submitted May 14, 2026. arxiv.org/...2606.12429
11. ^Meta Superintelligence Labs, "Introducing Muse Spark 1.1," July 9, 2026. ai.meta.com/...introducing-muse-spark-meta-model-api
12. ^Meta, "Meta AI Doesn't Just Think, It Acts," July 24, 2026. about.fb.com/...se-spark-doesnt-just-think-it-acts
13. ^Joelle Pineau, "Celebrating 10 years of FAIR: A decade of advancing the state-of-the-art through open research," Meta, November 30, 2023. ai.meta.com/...-year-anniversary-open-science-meta
14. ^Andrew Bosworth, "Building with AI across all of Meta," Meta, June 2, 2022. ai.meta.com/...building-with-ai-across-all-of-meta
15. ^Meta Platforms, Inc., "Q2 2025 Prepared Remarks," July 30, 2025. s21.q4cdn.com/...META-Q2-2025-Prepared-Remarks.pdf
16. ^Associated Press, "Meta invests $14.3B in AI firm Scale and recruits its CEO for 'superintelligence' team," June 13, 2025. apnews.com/...4b55aabf7ea018e38ffdccb66e37cf26
17. ^Reuters, "Meta plans fourth restructuring of AI efforts in six months, The Information reports," August 15, 2025. marketscreener.com/...ion-reports-ce7c51ded081f72d
18. ^CNBC, "Meta lays off 600 from 'bloated' AI unit as Wang cements leadership," October 22, 2025. cnbc.com/...meta-layoffs-ai
19. ^Reuters, "Meta transfers top engineers into new AI tooling team," April 10, 2026. finance.yahoo.com/...rs-top-engineers-ai-223108238
20. ^Mike Clark, "Privacy Matters: Meta's Generative AI Features," Meta, September 27, 2023, updated June 8, 2026. about.fb.com/...tters-metas-generative-ai-features
21. ^Meta Privacy Center, "How you can see or download the information about you from chats with AIs from Meta," accessed July 28, 2026. facebook.com/...see-or-download-ai-chat-info
22. ^Meta Privacy Center, "How Meta uses information for generative AI models and features," accessed July 28, 2026. facebook.com/...genai
23. ^Associated Press, "Meta says it will resume AI training with public content from European users," April 14, 2025. apnews.com/...c785dc3591ae3c49543c435fc15379fb
24. ^Meta, "Improving Your Recommendations on Our Apps With AI at Meta," October 1, 2025. about.fb.com/...-your-recommendations-apps-ai-meta
25. ^Drew Harwell and Gerrit De Vynck, "Meta AI chatbot is divulging users' most private searches," The Washington Post, June 13, 2025. washingtonpost.com/...meta-ai-privacy-users-chatbot
26. ^Meta, "Llama 4 Community License Agreement," April 5, 2025. github.com/...LICENSE
27. ^Jordan Maris, "Meta's LLaMa license is still not Open Source," Open Source Initiative, February 18, 2025. opensource.org/...license-is-still-not-open-source
28. ^Reuters, "Meta debuts Muse Spark 1.1 with preview open to developers," July 9, 2026. investing.com/...review-open-to-developers-4784384
29. ^Meta, "Introducing Muse Image: Image Generation Built for Your World," July 7, 2026, updated July 10, 2026. about.fb.com/...introducing-muse-image-meta-ai
30. ^Associated Press, "Amid criticism, Meta reins in new AI tool that automatically accessed public Instagram images," July 11, 2026. apnews.com/...acy-4df3bdb3fec6e046c6562accc2d270a5
31. ^Jeff Horwitz, Reuters, "Meta's AI rules have let bots hold 'sensual' chats with kids, offer false medical info," August 14, 2025. investing.com/...-offer-false-medical-info-4191596
32. ^Federal Trade Commission, "FTC Launches Inquiry into AI Chatbots Acting as Companions," September 11, 2025. ftc.gov/...s-inquiry-ai-chatbots-acting-companions
33. ^TechCrunch, "NYU 'Deep Learning' Professor LeCun Will Head Facebook's New Artificial Intelligence Lab," December 9, 2013. techcrunch.com/...rtificial-intelligence-lab-lecun
34. ^Ina Fried, Axios, "Meta revamps AI unit to get generative tech into products," February 27, 2023. axios.com/...meta-generative-ai-products
35. ^CNBC, "Facebook hires Jerome Pesenti as new VP of AI," January 23, 2018. cnbc.com/...k-hires-jerome-pesenti-as-new-vp-of-ai
36. ^CNBC, "Meta's head of AI research announces departure," April 1, 2025. cnbc.com/...ead-of-ai-research-announces-departure
37. ^TechCrunch, "Meta taps former Google DeepMind director to lead its AI research lab," May 8, 2025. techcrunch.com/...ctor-to-lead-its-ai-research-lab
38. ^CNBC, "Mark Zuckerberg announces creation of Meta Superintelligence Labs. Read the memo," June 30, 2025. cnbc.com/...a-superintelligence-labs-read-the-memo
39. ^TechCrunch, "Meta names Shengjia Zhao as chief scientist of AI superintelligence unit," July 25, 2025. techcrunch.com/...ist-of-ai-superintelligence-unit
40. ^CNBC, "Meta chief AI scientist Yann LeCun is leaving to create his own startup," November 19, 2025. cnbc.com/...ist-yann-lecun-is-leaving-the-company-
41. ^Mark Zuckerberg, "Personal Superintelligence," Meta, July 30, 2025. about.fb.com/...personal-superintelligence
42. ^Meta AI, "Introducing LLaMA: A foundational, 65-billion-parameter large language model," February 24, 2023. ai.meta.com/...large-language-model-llama-meta-ai
43. ^Meta, "Llama 2 model card," Hugging Face, July 18, 2023. huggingface.co/...Llama-2-70b-hf
44. ^Meta AI, "Introducing Meta Llama 3: The most capable openly available LLM to date," April 18, 2024. ai.meta.com/...meta-llama-3
45. ^Meta AI, "Introducing Llama 3.1: Our most capable models to date," July 23, 2024. ai.meta.com/...meta-llama-3-1
46. ^Meta AI, "Llama 3.2: Revolutionizing edge AI and vision with open, customizable models," September 25, 2024. ai.meta.com/...ect-2024-vision-edge-mobile-devices
47. ^Meta, "Llama 3.3 70B Instruct model card," Hugging Face, December 6, 2024. huggingface.co/...Llama-3.3-70B-Instruct
48. ^Meta AI, "The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation," April 5, 2025. ai.meta.com/...llama-4-multimodal-intelligence
49. ^Meta, "Llama-4-Maverick-17B-128E-Instruct model card," Hugging Face, April 2025. huggingface.co/...Llama-4-Maverick-17B-128E-Instruct
50. ^Meta, "Llama 3.2 Community License Agreement and Acceptable Use Policy," September 2024. github.com/...USE_POLICY.md
51. ^Meta, "Celebrating 1 Billion Downloads of Llama," March 18, 2025. about.fb.com/...ebrating-1-billion-downloads-llama
52. ^TechCrunch, "Meta says its Llama AI models have been downloaded 1.2B times," April 29, 2025. techcrunch.com/...-have-been-downloaded-1-2b-times
53. ^Meta AI, "Everything we announced at our first-ever LlamaCon," April 29, 2025. ai.meta.com/...llamacon-llama-news
54. ^CNBC, "Mark Zuckerberg says Meta AI has 1 billion monthly active users," May 28, 2025. cnbc.com/...berg-meta-ai-one-billion-monthly-users
55. ^Meta Platforms, Inc., "Meta Reports Second Quarter 2026 Results," July 29, 2026. prnewswire.com/...d-quarter-2026-results-302838214
56. ^Meta Platforms, Inc., "Q2 2026 Earnings Call Transcript," July 29, 2026. s21.q4cdn.com/...2026-Earnings-Call-Transcript.pdf
57. ^Meta Platforms, Inc., "Q1 2026 Earnings Call Transcript," April 29, 2026. s21.q4cdn.com/...2026-Earnings-Call-Transcript.pdf
58. ^Meta Platforms, Inc., "Meta Reports Fourth Quarter and Full Year 2025 Results," January 28, 2026. prnewswire.com/...full-year-2025-results-302673127
59. ^Meta Platforms, Inc., "Meta Reports First Quarter 2026 Results," April 29, 2026. investor.atmeta.com/...default
60. ^Meta AI, "MTIA v1: Meta's first-generation AI inference accelerator," May 18, 2023. ai.meta.com/...ining-inference-accelerator-AI-MTIA
61. ^Meta AI, "Our next generation Meta Training and Inference Accelerator," April 10, 2024. ai.meta.com/...ining-inference-accelerator-AI-MTIA
62. ^Yee Jiun Song, Andrew Tulloch, Harikrishna Reddy, CQ Tang and Vijay Thakkar, "Four MTIA Chips in Two Years: Scaling AI Experiences for Billions," Meta AI, March 11, 2026. ai.meta.com/...meta-mtia-scale-ai-chips-for-billions
63. ^Meta, "Expanding Meta's Custom Silicon to Power Our AI Workloads," March 11, 2026. about.fb.com/...-silicon-to-power-our-ai-workloads
64. ^Meta AI, "Introducing the AI Research SuperCluster: Meta's cutting-edge AI supercomputer for AI research," January 24, 2022. ai.meta.com/...ai-rsc
65. ^Meta, "Meta's Grand Teton brings NVIDIA Hopper to its data centers," OCP Global Summit, October 18, 2022. engineering.fb.com/...ocp-summit-2022-grand-teton
66. ^Meta, "Meta's open AI hardware vision," OCP Global Summit, October 15, 2024. engineering.fb.com/...metas-open-ai-hardware-vision
67. ^CNBC, "Mark Zuckerberg indicates Meta is spending billions of dollars on Nvidia AI chips," January 18, 2024. cnbc.com/...s-spending-billions-on-nvidia-ai-chips
68. ^Meta, "Meta Announces New Strategic Venture With BlackRock to Develop Data Center in El Paso," July 28, 2026. about.fb.com/...-to-develop-data-center-in-el-paso
69. ^SiliconANGLE, "Meta Platforms creates new organization to lead its AI infrastructure buildout," January 12, 2026. siliconangle.com/...ead-ai-infrastructure-buildout
70. ^Meta AI, "Introducing Segment Anything: Working toward the first foundation model for image segmentation," April 5, 2023. ai.meta.com/...foundation-model-image-segmentation
71. ^Meta AI, "SAM 2: Segment Anything in Images and Videos," July 29, 2024. ai.meta.com/...segment-anything-2
72. ^Meta, "New Segment Anything Models Make it Easier to Detect Objects and Create 3D Reconstructions," November 19, 2025. about.fb.com/...-objects-create-3d-reconstructions
73. ^Meta AI, "SAM 3.1: Faster and More Accessible Real-Time Video Detection and Tracking With Multiplexing and Global Reasoning," March 27, 2026. ai.meta.com/...segment-anything-model-3
74. ^Meta AI, "DINOv3: Self-supervised learning for vision at unprecedented scale," 2025. ai.meta.com/...dinov3-self-supervised-vision-model
75. ^Meta AI, "Introducing V-JEPA 2," June 11, 2025. ai.meta.com/...vjepa
76. ^Meta AI Research, "CWM: An Open-Weights LLM for Research on Code Generation with World Models," September 2025. ai.meta.com/...n-code-generation-with-world-models
77. ^Mark Zuckerberg, "Open Source AI Is the Path Forward," Meta, July 23, 2024. about.fb.com/...open-source-ai-is-the-path-forward
78. ^Meta, "Our Approach to Frontier AI," February 3, 2025. about.fb.com/...meta-approach-frontier-ai
79. ^Meta, "Advanced AI Scaling Framework, Version 2," April 7, 2026 (renamed from the Frontier AI Framework of February 3, 2025). ai.meta.com/...Meta_Advanced-AI-Scaling-Framework-v2
80. ^PyTorch Foundation, "PyTorch strengthens its governance by joining the Linux Foundation," September 12, 2022. pytorch.org/...PyTorchfoundation
81. ^Kadrey et al. v. Meta Platforms, Inc., No. 3:23-cv-03417, Order on Cross-Motions for Partial Summary Judgment, N.D. Cal., June 25, 2025. law.justia.com/...598
82. ^TorrentFreak, "Copyright Lawsuit Accuses Meta of Pirating Adult Films for AI Training," July 2025. torrentfreak.com/...ng-adult-films-for-ai-training
83. ^CNBC, "Meta says it won't sign Europe AI agreement, calling it an overreach that will stunt growth," July 18, 2025. cnbc.com/...meta-europe-ai-code
84. ^TechCrunch, "Meta pauses teen access to AI characters ahead of new version," January 23, 2026. techcrunch.com/...-characters-ahead-of-new-version
85. ^The Hill, "Meta adding AI chatbot safety features for teens," October 17, 2025. thehill.com/...72-meta-safety-features-ai-chatbots
86. ^Will Douglas Heaven, "Why Meta's latest large language model only survived three days online," MIT Technology Review, November 18, 2022. technologyreview.com/...d-three-days-gpt-3-science
87. ^Meta, "Introducing Vibes: A New Way to Discover and Create AI Videos," September 25, 2025. about.fb.com/...introducing-vibes-ai-videos
88. ^Meta, "Create Your Own Custom AI With AI Studio," July 29, 2024. about.fb.com/...-your-own-custom-ai-with-ai-studio
89. ^Meta, "Meta Ray-Ban Display: AI Glasses With an EMG Wristband," September 17, 2025. about.fb.com/...n-display-ai-glasses-emg-wristband
90. ^CNBC, "Ray-Ban maker EssilorLuxottica says it more than tripled Meta AI glasses sales in 2025," February 11, 2026. cnbc.com/...ttica-triples-sales-of-meta-ai-glasses
91. ^TechCrunch, "Meta's months-old AI unit is a soul-crushing gulag, say the engineers stuck inside it," June 12, 2026. techcrunch.com/...ay-the-engineers-stuck-inside-it
92. ^TechCrunch, "Meta exec denies the company artificially boosted Llama 4's benchmark scores," April 7, 2025. techcrunch.com/...oosted-llama-4s-benchmark-scores
93. ^Fast Company, "Yann LeCun: Meta 'fudged' on Llama 4 testing," January 2026. fastcompany.com/...n-meta-llama-4-model-zuckerberg
94. ^CNBC, "Meta acquires Singapore AI agent firm Manus, capping year of aggressive AI moves," December 30, 2025. cnbc.com/...m-manus-china-butterfly-effect-monicai
95. ^The Decoder, "Meta acqui-hires Dreamer's entire team to bolster its AI agent ambitions," March 2026. the-decoder.com/...-its-lagging-ai-agent-ambitions
96. ^The Next Platform, "Meta Buys Rivos To Accelerate Compute Engine Engineering," October 2, 2025. nextplatform.com/...ate-compute-engine-engineering
97. ^Daniel Bolya et al., "Perception Encoder: The best visual embeddings are not at the output of the network," arXiv:2504.13181, April 2025. arxiv.org/...2504.13181
98. ^Yung-Sung Chuang et al., "Meta CLIP 2: A Worldwide Scaling Recipe," arXiv:2507.22062, July 2025. arxiv.org/...2507.22062
99. ^Meta, "Meta Partners With AWS on Graviton Chips to Power Agentic AI," April 24, 2026. about.fb.com/...graviton-chips-to-power-agentic-ai
100. ^The Batch, DeepLearning.AI, "How Meta's LLaMA NLP Model Leaked," March 15, 2023. deeplearning.ai/...how-metas-llama-nlp-model-leaked
101. ^Fortune, "Meta's new AI chatbot BlenderBot 3 is giving painful, offensive responses," August 9, 2022. fortune.com/...onses-meta-blenderbot-3-chatbot-tay
102. ^Meta Engineering, "GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model," August 3, 2026. engineering.fb.com/...ommendation-foundation-model