digital-humans.org
Search

Meta AI – company AI strategy and products

A pillar analysis in Foundation Models – Vendors & Models. We examine the landscape, the trade-offs, and what matters in 2026.

Published Updated 15 min read

Meta AI is the artificial intelligence division of Meta Platforms, the parent company of Facebook, Instagram, WhatsApp and Reality Labs. Through 2025 it housed two distinct units – Fundamental AI Research (FAIR), founded in 2013 and led by Yann LeCun, and the applied product teams behind the Meta AI assistant inside the apps. In June 2025 Meta restructured the entire division into Meta Superintelligence Labs (MSL) under Alexandr Wang, the 28-year-old founder of Scale AI, after Meta took a 49% non-voting stake in Scale for $14.3 billion. LeCun resigned in November 2025 rather than report to Wang. As of May 2026 Meta is the third major frontier lab after OpenAI and Anthropic, with a fundamentally different distribution strategy: it releases its frontier models under an open-weights licence and lets the rest of the industry use them.

If you are evaluating whether to build on Llama, whether to take Meta seriously as a long-term foundation choice, or what Meta’s strategic position means for the open-weights ecosystem you depend on, the question is no longer “is Meta capable?” but “is the Meta of 2026 the same company that built the Llama of 2024?” The corporate restructuring matters as much as the model lineup. This article covers both.

Meta AI in one paragraph

The structural story over the past 18 months is sharper than the model lineup. Through early 2025 Meta’s AI organisation looked stable. FAIR had been running for 12 years under LeCun, a Turing Award winner widely seen as one of the three founding fathers of modern deep learning alongside Geoffrey Hinton and Yoshua Bengio. The Llama series, released by FAIR starting February 2023, had defined the open-weights frontier. The product applications – Meta AI assistant across Facebook, Instagram, WhatsApp, plus Ray-Ban Meta glasses – were maturing. Funding scale was at the level of the other frontier labs, with Meta’s 2025 capital expenditure on AI infrastructure projected at $66-72 billion.

Then the structure broke. In June 2025 Mark Zuckerberg paid $14.3 billion for a 49% non-voting stake in Scale AI and brought Alexandr Wang in to lead a new division called Meta Superintelligence Labs, alongside Nat Friedman, the former GitHub CEO. The new structure folded FAIR, the foundations team, the products team and a new “TBD Lab” (frontier-model effort) all under Wang. Zuckerberg called Wang “the most impressive founder of his generation” in an internal memo.

The disruption was substantial. In October 2025 Meta cut approximately 600 positions in MSL, largely from FAIR and the Products and Applied Research teams. In November 2025 Yann LeCun resigned after being asked to report to Wang, telling the Financial Times that Wang was “young and inexperienced” and predicting that “a lot of people have left, a lot of people who haven’t yet left will leave.” LeCun subsequently raised $1 billion to found AMI Labs in Paris in early 2026, drawing several FAIR researchers with him. By March 2026 Meta was reorganising MSL again, splitting it into four sub-groups – TBD Lab for foundation models, a product unit, an infrastructure team, and a reduced FAIR – and the company’s first closed-source frontier model, Muse Spark, had launched on 8 April 2026. The Llama-as-open-weights commitment is, as of May 2026, intact but no longer the only thing Meta ships.

Compute scale remains the differentiator nobody else in the open-weights world can match. Meta’s projected AI capital expenditure for 2026 sits in the $80-100 billion range, roughly comparable with Microsoft and Google. The combination of frontier-scale capex and the public open-weights commitment is unique. No other lab at this scale releases weights. We unpack the Microsoft/OpenAI infrastructure dependency separately in the OpenAI company profile and the comparison set in our Anthropic profile.

The Llama strategy explained

To understand why Meta releases frontier-class models open-weight, you have to read Mark Zuckerberg’s July 2024 letter on the topic. The core argument has three pieces.

First, commoditisation of the model layer benefits Meta more than any other major player. Meta does not sell AI as a service. It sells advertising on social products. If the underlying foundation model becomes a commodity – the way Linux became a commodity for server operating systems – Meta benefits because its competitors (Google, Microsoft, OpenAI, Anthropic) cannot use proprietary model access as a competitive moat against Meta’s product distribution. Linux did not win because Red Hat made it free; Linux won because the consortium of users who benefited from its commodity status was larger than the group of companies that wanted to keep operating systems proprietary. Zuckerberg’s bet is that the same dynamic will play out for foundation models.

Second, Meta cannot afford the platform-tax problem. Apple charges a 30% fee on the App Store. Google charges 15-30% on Play. If OpenAI or Google were to own the dominant foundation model that everyone, including Meta, has to build on, Meta would face an equivalent platform tax on its own AI features. By releasing Llama, Meta ensures it always has an alternative foundation model it controls. The release is not altruism; it is insurance.

Third, the research advantages of open release are real. FAIR’s open-publication tradition (PyTorch, OPT, the Llama papers, SAM, ImageBind, AudioCraft, SeamlessM4T) has historically attracted top research talent. In his July 2024 letter, Zuckerberg argued that Meta benefits from improvements the broader community makes to Llama-class models. Whether this still holds in 2026 after the FAIR cuts and LeCun’s departure is open.

What Llama is not, despite the marketing, is genuinely open source. The Llama Community License is source-available with two material restrictions. The first is the 700-million-MAU clause. The licence text reads: “If, on the Llama 4 version release date, the monthly active users of the products or services made available by or for Licensee, or Licensee’s affiliates, is greater than 700 million monthly active users in the preceding calendar month, you must request a license from Meta, which Meta may grant to you in its sole discretion.” Practically, this excludes Apple, Google, Microsoft, ByteDance, Tencent and Amazon at the corporate-entity level – not at the individual-product level. The clause uses the term “Licensee’s affiliates,” which means the user count is summed across the entire corporate group. The second restriction prohibits using Llama outputs to improve any competing foundation model. Both clauses are the reason the Open Source Initiative has never recognised the Llama licence as open source.

For the 99.9% of organisations under 700M MAU, neither restriction binds. For the largest tech companies, Llama is functionally unusable without a separate commercial deal with Meta.

The Llama lineup as of May 2026

The current family was released on 5 April 2025 and remains the production lineup. Three models: Scout, Maverick, Behemoth.

Llama 4 Scout is the efficiency model. 17 billion active parameters across 16 experts, 109 billion total parameters via mixture-of-experts (MoE) architecture. Native multimodality (text plus images). The headline feature is a 10-million-token context window, the longest of any publicly available model as of mid-2026. Runs on a single H100 GPU. Used for document analysis, codebase processing, and long-context retrieval workflows.

Llama 4 Maverick is the performance model. 17 billion active parameters across 128 experts, 400 billion total parameters. 1-million-token context. Native multimodal. Meta’s published benchmarks show Maverick beating GPT-4o and Gemini 2.0 Flash on standard evaluations, with an experimental chat variant scoring an ELO of 1417 on LMSys Arena at launch. Maverick currently runs at approximately $0.49 per million tokens through API providers like Fireworks, Together and OpenRouter, against $2.50 per million for GPT-5.4 input – roughly one-fifth the cost.

Llama 4 Behemoth is the teacher model. 288 billion active parameters across 16 experts, nearly 2 trillion total parameters. Designed not for production deployment but to distill knowledge into Scout and Maverick during training. Meta’s internal benchmarks position Behemoth above GPT-4.5, Claude Sonnet 3.7 and Gemini 2.0 Pro on STEM-focused tests including MATH-500 and GPQA Diamond, though Gemini 2.5 Pro outperformed it on some evaluations. Behemoth has not been publicly released as open weights and may never be; Meta’s communications describe it as “in training” with no committed timeline.

The MoE architecture is the key story under the hood. Maverick has 400 billion total parameters but activates only 17 billion per token, which means inference cost and latency are closer to a 17B-parameter dense model while the knowledge base is closer to a 400B model. This is the same architectural pattern as DeepSeek V3 and the more recent OpenAI o-series. It is now the default approach for frontier-class open-weights releases.

Llama versus the alternatives

Where Llama 4 leads:

  • Cost per token via API. Roughly 1/5th to 1/10th the cost of frontier proprietary models for comparable quality on many tasks.
  • Self-hosting. Open weights mean you can run inference on your own hardware, control the data path, fine-tune on proprietary data without exposing it, and avoid per-call billing entirely.
  • Long context. Scout’s 10M-token window is structurally larger than anything proprietary models offer. For document-heavy or codebase-scale workloads this matters.
  • Customisation. Full fine-tuning, LoRA adapters, quantisation, distillation – all open. With proprietary models you get the fine-tuning APIs the vendor chooses to expose.

Where Llama 4 trails:

  • Frontier reasoning and computer use. Maverick is competitive with GPT-4o but falls short of GPT-5.5 on complex reasoning and agentic computer-use tasks. The gap on the hardest benchmarks (FrontierMath, Terminal-Bench, SWE-Bench Verified) is real.
  • Instruction following on edge cases. Proprietary models with extensive RLHF and post-training generally produce cleaner output on adversarial prompts, multi-turn instruction following, and refusal behaviour.
  • Multimodal depth. Native multimodality is in Llama 4, but the visual reasoning quality lags GPT-5.5 and Gemini 3.1.
  • Tooling and ecosystem. Function calling, structured outputs, embeddings, fine-tuning APIs, error states – the integrated experience of OpenAI and Anthropic APIs is substantially smoother than the hosted Llama options.

For full comparison we cover Claude in the Claude model family explainer and Gemini in the Gemini family explainer. The deeper Llama-specific architecture analysis lives in the Llama deep dive, and self-hosting choices are covered in Llama 3 vs Llama 4 for local deployment.

The other Meta AI products

Llama is the foundation, but Meta AI ships several other things worth knowing about.

Meta AI assistant is the consumer-facing chatbot integrated into WhatsApp, Instagram, Facebook and Messenger, reportedly reaching over 600 million monthly active users by mid-2025 through the in-app integration. It runs on Llama 4 with retrieval augmentation. Most users do not realise they are using an AI assistant in the same product family as ChatGPT – distribution by integration rather than by separate app.

Ray-Ban Meta glasses are the embodiment hardware. The second-generation models shipped in 2024 with live AI, multimodal vision and audio. Meta has positioned these as the form factor where AI becomes ambient rather than screen-bound. We cover the category in detail in our AI glasses pillar.

Reality Labs continues the longer-horizon bet on AR/VR/MR as the next computing surface, with Meta’s Orion prototype announced in September 2024 representing the first credible consumer AR glasses target. The AI strategy and the embodiment strategy are increasingly intertwined: glasses need on-device inference, which needs efficient small models like Llama 4 Scout, which justifies the open-weights strategy as the way to get there.

AudioCraft, MusicGen, SeamlessM4T and the rest of the FAIR-published model line have continued shipping under open licences. SeamlessM4T (speech-to-speech translation across 100 languages) and ImageBind (six-modality joint embeddings) are the standouts. Whether this research output continues at the same cadence after the FAIR cuts and LeCun’s departure is the open question.

Muse Spark, launched 8 April 2026, is the first closed-source frontier model from Meta Superintelligence Labs. This is the structural signal. The open-weights commitment is intact for Llama, but Meta is no longer exclusively an open-weights lab. The MSL frontier work is proprietary.

For builders: when Llama is the right choice

Three decisions cover most cases.

If you need frontier-class model quality at the lowest per-token cost, Llama 4 Maverick via API is the credible choice. Run it through Fireworks, Together, OpenRouter or Replicate. Expect roughly 80% of GPT-5.5 quality at roughly 20% of the cost. For high-volume workloads where cost is the binding constraint (chatbots at scale, content generation, summarisation pipelines), the math favours Llama. We cover the API-routing patterns in the agentic AI silo.

If you need to control the data path – health-care, finance, legal, government, anything with data-residency or audit requirements – self-hosted Llama is often the only configuration that survives compliance review. Open weights mean the model never sees the public internet during inference, and your inputs never leave your infrastructure. This is the genuinely defensible reason for choosing Llama over closed models, and it is where the open-weights commitment converts most directly to enterprise value.

If you need maximum customisation – fine-tuning on proprietary data, quantisation for edge deployment, custom inference engines – Llama is the only frontier-class option. Full fine-tuning, LoRA, QLoRA, GGUF for llama.cpp, MLX for Apple Silicon – the entire open ecosystem builds on Llama because the weights are downloadable.

The cases where Llama is the wrong choice are equally clear. Frontier reasoning at the limit (mathematical proofs, complex multi-step agentic tasks): use GPT-5.5 or Claude Opus 4.7. Production chat with extensive RLHF polish on adversarial inputs: GPT-5.5. Multimodal depth (video understanding, complex visual reasoning): Gemini 3.1 Pro. Multilingual coverage outside English-Chinese-Spanish at frontier quality: Gemini or Claude. The right architecture for most production workloads is a routing layer that uses Llama for cost-controlled bulk work and a proprietary frontier model for the hardest 10-15% of queries.

The strategic risks

Three risks deserve naming explicitly.

Llama licence stability. The 700M MAU clause has not changed across Llama 2, Llama 3, Llama 3.1, Llama 3.2 and Llama 4. The Acceptable Use Policy can change unilaterally without changing the licence itself. Meta could, in principle, tighten either at any future release. The track record so far is conservative; the strategic incentive to do so does not currently exist. But the risk is non-zero, and it belongs in any commercial risk register that bets on Llama as a long-term foundation choice.

US-China export and trade policy. Llama weights are subject to US export controls. The Acceptable Use Policy explicitly prohibits use in support of certain restricted activities, and US trade policy on AI exports has tightened substantially through 2024-2026. Builders deploying Llama in jurisdictions or for use cases that touch US export-controlled areas should treat this as a compliance question, not a technical one.

Meta’s future as a frontier lab. The departure of LeCun and the dissolution of FAIR’s autonomous research model are real shifts. Whether Meta Superintelligence Labs under Wang can maintain the open-weights commitment at frontier capability is the question that matters most for builders depending on Llama as a long-term substrate. The launch of Muse Spark as closed-source in April 2026 is the signal in one direction; the continued release of Llama 4 across Scout, Maverick and the eventual Behemoth is the signal in the other. The bet to make is on Meta’s economic incentive to keep open-weights frontier alive (the platform-tax argument), not on the cultural commitment to research openness that defined FAIR.

If you are building on Llama for the next five years, plan for the licence to remain stable, the model lineup to keep improving roughly annually, and the open-weights commitment to hold at the frontier minus one generation. Anything beyond that is speculation.

Frequently asked questions

What is Meta AI?

Meta AI is the artificial intelligence division of Meta Platforms (Facebook, Instagram, WhatsApp, Reality Labs). As of May 2026 it is organised as Meta Superintelligence Labs (MSL), led by Alexandr Wang. It develops the Llama family of foundation models, the Meta AI assistant deployed across the company’s apps, the AI features inside Ray-Ban Meta glasses, and research outputs from the (now reduced) FAIR research lab.

Who runs Meta AI?

Alexandr Wang, the 28-year-old founder of Scale AI, has been chief AI officer at Meta since June 2025, leading Meta Superintelligence Labs. Shengjia Zhao, a former lead scientist at OpenAI and co-creator of ChatGPT, serves as chief scientist of Superintelligence Labs. Yann LeCun, the previous chief AI scientist, departed Meta in November 2025.

What is Llama and how does it differ from ChatGPT?

Llama is Meta’s family of open-weights foundation models. ChatGPT is OpenAI’s closed-source chat product. The functional difference: ChatGPT runs only through OpenAI’s API and consumer interface, with full vendor control. Llama weights can be downloaded, fine-tuned, self-hosted, and run on your own infrastructure. The Llama Community License restricts companies with more than 700 million monthly active users from using Llama without Meta’s permission, which is why it is not formally open source.

Is Llama free for commercial use?

For organisations under 700 million monthly active users (essentially all but a handful of the largest tech companies), yes, with two restrictions. You must include the Llama attribution notice in distributed copies, and you cannot use Llama outputs to improve other foundation models. For companies over 700M MAU, a separate licence agreement with Meta is required, granted at Meta’s sole discretion.

What is the difference between Llama 4 Scout, Maverick and Behemoth?

Scout (17B active / 109B total parameters, 10M-token context) is the efficiency model, designed for single-GPU deployment and long-context workflows. Maverick (17B active / 400B total parameters, 1M-token context) is the performance model, designed for general production workloads. Behemoth (288B active / 2T total parameters) is the teacher model used to train Scout and Maverick via distillation; it has not been publicly released.

Why does Meta release Llama open-weight?

Three reasons, per Mark Zuckerberg’s July 2024 letter on open-source AI:

  • Commoditising the model layer prevents competitors from gaining a foundation-model moat over Meta.
  • Avoiding a future “platform tax” if a proprietary foundation model becomes dominant.
  • Research advantages from open community contributions.

The deeper strategic logic is that Meta’s revenue comes from products (advertising), not from AI, so making foundation models a commodity benefits Meta without cannibalising its business model.

How does Llama 4 compare to GPT-5.5?

Llama 4 Maverick lags GPT-5.5 on complex reasoning, agentic computer-use tasks, and the hardest STEM benchmarks. It is competitive or better on cost (roughly 1/5th the price), context window (Scout’s 10M tokens), customisation (full fine-tuning available), and self-hosting flexibility. For most production workloads the right architecture is to route to Llama for cost-controlled bulk work and to a frontier proprietary model for the most demanding 10-15% of queries.

Is Meta still committed to open-source AI?

For Llama specifically, yes, through 2026. The launch of Muse Spark as Meta’s first closed-source frontier model in April 2026 signals that Meta is no longer exclusively an open-weights lab, but the Llama commitment is intact. The longer-term commitment depends on whether the economic logic Zuckerberg laid out in 2024 continues to hold, particularly the platform-tax argument against proprietary foundation models.

What is Meta Superintelligence Labs?

Meta Superintelligence Labs (MSL) is the AI division Meta created in June 2025 under Alexandr Wang. It absorbed FAIR (the prior research lab), the foundations team, the applied product team, and a new TBD Lab focused on frontier model development. The restructure followed Meta’s $14.3 billion investment for a 49% non-voting stake in Scale AI. Approximately 600 positions were cut from MSL in October 2025. The unit was reorganised again in early 2026 into four sub-groups: TBD Lab, Products, Infrastructure, and a reduced FAIR.

What happened to Yann LeCun?

Yann LeCun served as Meta’s chief AI scientist for 12 years before resigning in November 2025 after being asked to report to Alexandr Wang. He told the Financial Times that Wang was “young and inexperienced” and warned that “a lot of people have left, a lot of people who haven’t yet left will leave.” LeCun raised $1 billion to found AMI Labs in Paris in early 2026, drawing several former FAIR researchers with him.

Should my company build on Llama?

If your priorities are cost-controlled inference, data-path control (regulated industries), or maximum customisation, yes. Run Maverick via Fireworks, Together or OpenRouter, or self-host Scout for compliance-sensitive workloads. If you need frontier reasoning quality at the limit, complex agentic tasks, or polished instruction-following on adversarial prompts, use GPT-5.5 or Claude Opus 4.7 instead. The most resilient production architecture routes different tasks to different models rather than committing to one.

Where can I download Llama 4?

Llama 4 Scout and Maverick are available at llama.com (registration required), Hugging Face (under Meta’s organisation), and through cloud-provider integrations including AWS SageMaker, Azure ML and Google Vertex AI. API access without self-hosting is available through Fireworks AI, Together AI, OpenRouter, Replicate, and Groq. Behemoth has not been released and may not be.