Meta AI

Meta AI is the artificial intelligence organisation of Meta Platforms, the company formerly known as Facebook. It began in December 2013 as Facebook AI Research, a small laboratory led by Yann LeCun, and grew into a division that now supplies models to billions of users and publishes weights that much of the open-source ecosystem builds on 1.

The organisation's most consequential decision was to release the Llama models for others to download and run. That choice made Meta the reference point for open-weight AI, put its architecture at the centre of a global research community, and created a competitive dynamic that closed laboratories have had to answer 2.

Meta AI is also the least conventional company in this volume. It sells no model access at list prices, operates no standalone AI cloud, and measures success partly by the adoption of models it gives away. Its commercial logic runs through advertising, devices and the assistant embedded in Meta's applications 1.

Facebook AI Research opened with a deliberate academic model: publish, hire on research reputation, and give scientists freedom to work on problems without a product deadline. LeCun, a pioneer of convolutional networks, led it, and the laboratory recruited heavily from universities in Europe and North America 1.

The bet produced results. FAIR researchers contributed to foundational work on contrastive learning, self-supervised representations, retrieval and efficient training, and Meta released PyTorch, the framework that became the default for academic machine learning. Publishing was not incidental; it was how the laboratory attracted people who would otherwise have stayed in academia 1.

FAIR's culture sat awkwardly beside a company whose products depended on ranking and advertising. For years the research group and the applied AI teams operated separately, with different metrics and different tolerances for risk. The 2023 reorganisation, which created a single generative AI group, was an attempt to close that gap under competitive pressure 1.

Meta AI is now larger and more product-driven than FAIR was. The research brand survives, but the organisation's output is judged by whether it improves the assistant, the advertising system and the devices.

The first Llama model, released to researchers in February 2023, was a family of dense models from 7 to 65 billion parameters trained on trillions of tokens. Its weights were not meant to be public; they leaked within days, and the episode demonstrated both demand for open models and the difficulty of controlling them 1.

Llama 2, in July 2023, was the calculated response. Meta published three sizes, permitted commercial use under an acceptable-use policy, and partnered with Microsoft on distribution. The licence was not an open-source licence: it imposed limits, including a requirement of a separate agreement for products with more than 700 million monthly users, and it restricted certain uses 3.

Later generations widened the family and the ambition. Llama 3, in April 2024, moved to a new tokeniser and better data; Llama 3.1, in July 2024, included a 405-billion-parameter model, the largest openly downloadable model at the time; Llama 3.2 added vision and small on-device models; Llama 4, in April 2025, moved to mixture-of-experts designs with far more total than active parameters 4.

The strategy has been influential beyond Meta. Start-ups built products on Llama because it was free to run and could be fine-tuned; cloud providers offered it because customers asked; governments and universities used it because it did not require a foreign API. Meta gained an ecosystem it did not have to pay for 2.

Meta AI, the consumer assistant, launched in September 2023 inside WhatsApp, Messenger and Instagram, and later as a standalone application. It answers questions, generates images, and searches the web, and it is positioned less as a destination than as a feature of applications people already open every day 1.

The distribution is the product. Meta's applications reach more than three billion people, and an assistant embedded in a messaging thread can be used by people who would never install a separate chatbot. That reach is why Meta can compete in consumer AI without charging for model access.

The Ray-Ban Meta glasses, released in October 2023, added a camera and a voice interface, and a multimodal update in 2024 let the assistant describe what the wearer was looking at. The device is the clearest example of Meta's bet that AI will be consumed through hardware that is worn rather than opened 1.

The assistant has also been the company's main route to training data and feedback at scale. Every conversation, correction and image request is a signal, subject to the company's privacy terms, and the volume is far larger than any laboratory's paid user base.

In June 2025 Meta created Meta Superintelligence Labs, a new organisation with a mandate to pursue frontier research, co-led by Alexandr Wang, the chief executive of Scale AI, and Nat Friedman, the former chief executive of GitHub. Wang became Meta's chief AI officer 5.

The announcement came with a transaction: Meta agreed to invest about 14.3 billion dollars in Scale AI for close to half the company, a deal that valued the data-labelling business far above its previous private valuation and gave Meta access to its data operations 5.

The laboratory was also built by hiring. Meta approached researchers at OpenAI, Google DeepMind and Anthropic with offers reported in the tens and hundreds of millions of dollars, and several accepted. The recruiting campaign became a public story about the price of frontier talent, and it reshaped the labour market for AI researchers 1.

The creation of a separate superintelligence unit was an admission that the existing organisation had not kept pace. Meta's models were competitive but rarely leading, and its research output, while strong, had not produced a frontier system. The new group was given a broader mandate and a larger budget, and told to build one 5.

Meta's advantage is capital. Its advertising business generates tens of billions of dollars of free cash flow a year, and the company has directed a growing share of it into data centres, accelerators and power. Capital spending passed 39 billion dollars in 2024 and was guided above 60 billion for 2025, most of it for AI infrastructure 1.

The company has assembled clusters of hundreds of thousands of GPUs and has designed its own server and networking systems to run them. It has also invested in custom silicon and in power generation, including long-term contracts for nuclear and renewable electricity to supply data centres whose demand is measured in gigawatts 1.

Unlike OpenAI or Anthropic, Meta does not need to earn back its AI spending through model sales. The investment is justified by better recommendations, better advertising tools, and the possibility that the assistant becomes a gateway to Meta's products. That makes the spending easier to sustain and harder to evaluate 2.

The commercial returns are indirect and difficult to isolate. Meta reports AI-driven improvements in ad performance and engagement, but it does not break out AI revenue, and the company's own guidance treats the spending as a strategic necessity rather than a project with a payback period 1.

The Llama 4 launch in April 2025 produced the most concrete dispute. Meta submitted an experimental, unreleased version of Llama 4 Maverick to LM Arena, a crowd-sourced leaderboard where users compare anonymous model outputs, and the version scored well. After researchers noticed that the publicly downloadable Maverick behaved differently and ranked lower, the leaderboard's maintainers apologised, changed their policies to require disclosure of tested variants, and re-scored the public model 6.

Meta denied that it had tuned models for the benchmark. The episode nonetheless became the standard example of a broader problem: leaderboards measure what a laboratory chooses to submit, and a favourable position can be worth more than an honest one. It also illustrated the difficulty of comparing a hosted model with a downloadable one, which may be quantised or served with different settings 6.

The open-weights claim has been contested since Llama 2. Meta describes Llama as open source, but its licence restricts commercial use above a user threshold, imposes acceptable-use terms, and does not disclose training data. The Open Source Initiative's definition, published in October 2024, requires that a system's data information, code and weights be available in ways that permit study, modification and sharing for any purpose; by that standard, Llama is open-weight, not open-source 7.

Meta's answer is that the distinction is academic. Its licence permits most uses, including commercial ones for smaller companies, and the ability to download and run a model is what matters to developers. Critics reply that a licence which can be withdrawn or conditioned is a different kind of artefact from software under an OSI-approved licence 2 7.

Whether open weights make AI safer is the deepest disagreement in the field, and Meta is its most interested party. Meta argues that open models are more scrutinised, more widely deployed and harder to concentrate in a few hands, so that releasing them reduces risk over time. Critics argue that weights cannot be recalled, that fine-tuning can remove safety training, and that the diffusion of capability outpaces the diffusion of safeguards 2.

A second disagreement concerns the value of the spending. Meta's investors have generally accepted the capital expenditure as necessary; sceptics note that the company has not demonstrated a direct return, that its models are not the revenue leaders, and that the spending is justified largely by fear of missing the next platform 1.

A third concerns the superintelligence laboratory itself. Creating a separate unit with a broad mandate and large offers to individual researchers is either a serious commitment to frontier work or an expensive signal, depending on what it produces. The record so far is a hiring campaign and an organisational chart, not a model 5.

A fourth question is what openness does to regulation. Meta argues that open weights let regulators and independent researchers inspect systems that would otherwise be opaque, and that closed models are harder to audit. The counter-argument is that a downloadable model cannot be governed by the rules that apply to an API: there is no usage policy to enforce, no rate limit to adjust, and no way to withdraw a capability once the weights have been released 2.

Meta AI is the most consequential non-American-style competitor in the open-weight market — an American company that behaves, in model distribution, like a public utility. Llama derivatives account for a large share of the models downloaded and fine-tuned worldwide, and the family set the price of capable inference for everyone 4.

Its position differs from that of OpenAI or Anthropic. It does not need to win on model quality to benefit from AI, because its distribution and its advertising business give it other routes to return. That lowers the stakes of any single release and raises the stakes of the ecosystem 1.

Competition is now global. Chinese open-weight families such as Qwen and DeepSeek have become competitive with Llama on public benchmarks, and several are released under more permissive licences. Meta's ecosystem advantage is real, but it is no longer unique, and the company's decision to keep publishing weights depends on staying close enough to the frontier for its releases to matter.

For the encyclopedia, Meta AI is the test case for the argument that openness is a competitive strategy rather than a philosophy. The company gives away what its rivals sell, and it does so because the gift strengthens the platform on which its actual business runs 2.