Meta AI / FAIR

Meta AI is the artificial intelligence research and product division of Meta Platforms, headquartered in Menlo Park, California. It develops the Llama family of open-weights large-language models and, since 2025, the closed-weights frontier line under Meta Superintelligence Labs (MSL), the restructured organization led by Chief AI Officer Alexandr Wang. As of April 2026, Meta is in the middle of a strategic pivot from a research-and-open-weights orientation toward a frontier-capability and closed-source posture, anchored by the April 2026 launch of Muse Spark.

At a glance

• Founded: Facebook AI Research (FAIR) launched in 2013; Meta Superintelligence Labs (MSL) created in June 2025.
• Status: Subsidiary of Meta Platforms, Inc. Not independently capitalized.
• Funding: Operates within Meta's R&D budget. Parent company Meta Platforms has a market capitalization of approximately $1.5 trillion as of April 2026 and committed an additional $14.3 billion for a 49 percent stake in Scale AI in June 2025 alongside the MSL launch.
• Chief AI Officer: Alexandr Wang (since June 2025)
• Other notable leadership: Mark Zuckerberg (CEO, Meta Platforms; sets AI direction), Joelle Pineau (former VP of AI Research; departed in 2025), Nat Friedman (joined MSL leadership in 2025)
• Open weights: Mixed. The Llama family through Llama 4 is open weights. Muse Spark and subsequent MSL releases are closed.
• Flagship models: Muse Spark (April 2026), Llama 4 (April 2025), Llama 4 Behemoth (August 2025)

Origins

Meta's AI research history begins in 2013 with the founding of Facebook AI Research (FAIR), led from inception by Yann LeCun, the deep-learning researcher who would later share the 2018 Turing Award. FAIR established Facebook (later Meta) as a leading academic-research-style AI lab through a decade of open publications, the PyTorch deep-learning framework, and contributions to computer vision, natural language processing, and reinforcement learning.

The Llama family of open-weights large-language models, beginning with Llama 1 in February 2023, made Meta the leading source of open-weights frontier capability. Llama 2 (July 2023), Llama 3 (April 2024), and Llama 4 (April 2025) extended the family. Llama 4's release was followed by reports that Meta researchers had used different model versions on different benchmarks, which Yann LeCun later characterized in public as having "fudged a little bit." The episode contributed to internal frustration with the Llama line's competitive position.

In June 2025, Mark Zuckerberg announced a major restructuring of Meta's AI organization. The company invested $14.3 billion for a 49 percent stake in Scale AI, the data-labeling and model-evaluation company, and brought Scale CEO Alexandr Wang to Meta as its first-ever Chief AI Officer. The restructuring established Meta Superintelligence Labs (MSL) as the consolidated AI organization, absorbing FAIR and the Llama-development teams under unified leadership.

In November 2025, Yann LeCun departed Meta after twelve years to found Advanced Machine Intelligence Labs, an Insurgent venture pursuing video- and spatial-data architectures (V-JEPA) as alternatives to large-language models. LeCun's departure, the Scale AI deal, and Wang's elevation to Chief AI Officer together represent the largest organizational change in Meta's AI history.

The first MSL closed-source model, Muse Spark, launched on April 8, 2026.

Mission and strategy

Meta's stated AI mission is "to give people the power to build community and bring the world closer together" through advanced AI capabilities, applied across Meta's product surface (Facebook, Instagram, WhatsApp, Threads, Quest VR, and Ray-Ban smart glasses). The MSL framing extends this with an explicit "superintelligence" research goal, in language closer to OpenAI's and Anthropic's AGI missions than to FAIR's earlier academic-research framing.

The 2025 restructuring marks a strategic pivot. The old Meta AI strategy emphasized open-weights model releases (Llama family) as developer-ecosystem leverage and FAIR's academic publication record as research-credibility infrastructure. The new MSL strategy emphasizes frontier-capability competition with OpenAI, Anthropic, and Google DeepMind, with closed-source releases when frontier-tier results require it and open-source releases retained where they serve developer adoption.

The strategic premise underlying the pivot is that frontier capability has become the binding constraint for Meta's product strategy across its consumer apps and AR/VR platforms, and that the previous academic-research-and-open-weights model could not deliver competitive frontier capability quickly enough. The Scale AI investment provides data infrastructure; the Wang appointment provides commercial-product orientation; the closed-source model line provides flexibility to deploy frontier capability without immediate weights release.

Distribution is Meta's structural advantage. Llama and MSL models are deployed across Facebook, Instagram, WhatsApp, Threads, Meta AI assistant, Ray-Ban Meta smart glasses, and Quest VR headsets, with billions of monthly active users across the family.

Models and products

• Llama family. Llama 1 (February 2023), Llama 2 (July 2023), Llama 3 (April 2024), and Llama 4 (April 2025) constitute the open-weights line. Llama 4 includes Maverick (109B MoE, 17B active), Scout (smaller MoE), and the closed Behemoth variant (2T MoE, 288B active, August 2025). The Llama family has been the most-downloaded open-weights frontier model line on Hugging Face through most of 2024 and 2025.
• Muse Spark. First closed-source MSL frontier model, launched April 8, 2026. Positioned for advanced reasoning across science, math, and health, distinct from the Llama line's general-purpose framing.
• Code Llama, Llama Guard, Llama Stack. Specialized open-weights derivatives addressing coding (Code Llama), safety classification (Llama Guard), and developer infrastructure (Llama Stack).
• Meta AI assistant. Consumer assistant integrated into WhatsApp, Instagram, Facebook, Messenger, and the Ray-Ban Meta smart glasses. Uses Meta's frontier models depending on context.
• Segment Anything. Open-source computer vision model family for image and video segmentation, an early FAIR research-and-release contribution.
• PyTorch. Originally an FAIR project, now under the Linux Foundation; remains the dominant deep-learning framework globally.

Distribution is unique among the Frontier and Incumbent labs: Meta deploys models across consumer apps with billions of users (Facebook, Instagram, WhatsApp, Threads), AR and VR hardware (Quest, Ray-Ban Meta), and through the open-weights ecosystem on Hugging Face for Llama variants.

Benchmarks and standing

Llama 4's benchmark position has been contested. Meta's published benchmark results for Llama 4 Maverick and Scout were later acknowledged to have used model variants different from the publicly released versions, with Yann LeCun stating publicly that Meta had "fudged a little bit" on the benchmark numbers. Independent third-party evaluations have placed Llama 4 below the leading closed-source frontier models on most general-purpose benchmarks while remaining the strongest open-weights frontier option for many use cases.

Muse Spark's benchmark position is not yet established at scale; the April 8, 2026 release is recent enough that comparative evaluations are still in progress at the major leaderboards.

The standing of Llama as an open-weights line is structurally distinctive. Llama models are the most-deployed open-weights frontier models in many enterprise and developer contexts, particularly in regulated environments where on-premises inference is required. The competitive pressure on this position has come from Mistral AI's Mixtral and Mistral Large series, DeepSeek's V3 and R1, and increasingly from Alibaba's Qwen line.

Leadership

As of April 2026, Meta AI's senior leadership includes:

• Mark Zuckerberg, CEO of Meta Platforms. Sets overall AI direction and committed personally to the 2025 restructuring including the Scale AI investment and Wang's appointment.
• Alexandr Wang, Chief AI Officer (since June 2025). Co-founder and former CEO of Scale AI. Leads Meta Superintelligence Labs and consolidates AI strategy across Meta's product surface. Zuckerberg has called Wang "the most impressive founder of his generation."
• Nat Friedman, MSL leadership (joined 2025). Former GitHub CEO; focuses on AI infrastructure and product integration.
• Daniel Gross, MSL leadership (joined 2025). Investor and former Y Combinator partner; focuses on talent and applied research.

The departures cohort is unusually significant. Yann LeCun departed in November 2025 after twelve years and founded AMI Labs. Joelle Pineau departed in 2025 from her Vice President of AI Research role. Several senior FAIR researchers left during the 2025 restructuring, with Meta also reportedly hiring five Thinking Machines Lab founders in concurrent talent moves.

Funding and backers

Meta AI has not raised independent capital. The organization operates within Meta Platforms' R&D budget. Meta Platforms has a market capitalization of approximately $1.5 trillion as of April 2026, making the parent company among the most valuable in the world.

The most significant capital event in Meta's AI history is the June 2025 Scale AI transaction. Meta paid approximately $14.3 billion for a 49 percent stake in Scale AI, with the deal also bringing Scale CEO Alexandr Wang to Meta as Chief AI Officer. The transaction was structured to align Scale's data-labeling capacity with Meta's training pipeline while preserving Scale's independent corporate identity.

Meta's broader AI capital expenditure has been substantial. Public guidance through 2025 and 2026 has indicated tens of billions of dollars in annual capex on AI infrastructure, including new data centers, GPU procurement, and the talent acquisition cost of the MSL leadership build.

Industry position

Meta AI occupies a structurally distinctive position among Incumbent AI labs. The combination of consumer-product distribution at billions-of-users scale, the historical open-weights commitment via Llama, and the 2025 strategic pivot toward frontier-capability competition produces a profile unlike Apple's, Microsoft's, or Amazon's AI organizations. The Llama family's ecosystem leverage has been Meta's primary AI moat through 2024 and 2025; the Muse Spark closed-source release marks a deliberate complement rather than a replacement.

Strategic risks identified in industry coverage include the integration challenge of merging FAIR's research-academic culture with Scale's commercial-execution culture under Wang's leadership, the loss of Yann LeCun's research-credibility signal, the open question of whether Meta can close the capability gap to OpenAI, Anthropic, and Google DeepMind given the comparatively late start on the closed-source frontier line, and reputational damage from the Llama 4 benchmark disclosure.

The strategic strengths are equally distinct. Distribution at consumer scale is unmatched outside Google. The Scale AI integration provides differentiated data infrastructure. The Llama ecosystem maintains developer reach. The combined Meta and Scale technology and talent footprint is one of the largest single AI capability concentrations globally.

Competitive landscape

Meta AI competes with several Frontier and Incumbent labs:

• OpenAI. Direct frontier-capability competitor, particularly for consumer AI and product-integration use cases. Meta's Muse Spark is positioned as a frontier-tier alternative to GPT-5.5.
• Anthropic. Less direct given Anthropic's enterprise focus, but Anthropic is the principal alternative for enterprise customers evaluating frontier capability.
• Google DeepMind. Closest peer among Incumbent AI labs by structural shape (subsidiary of large public-tech parent, multiple consumer-product distribution surfaces). Gemini in Google's apps competes with Meta AI in Meta's apps.
• Mistral AI. Open-weights frontier competitor; Mistral and Mixtral have grown developer share previously held by Llama.
• DeepSeek and Alibaba Qwen. Open-weights pressure from Chinese frontier labs has accelerated; both produce competitive open-weights models at lower cost than the Llama line.
• Microsoft AI and Apple. Other Incumbent AI organizations with substantial product distribution; competitive pressure is more about consumer-AI differentiation than direct model-capability head-to-heads.

Outlook

Several open questions affect Meta AI's trajectory in 2026 and 2027:

• The capability profile and adoption of Muse Spark and any subsequent MSL releases relative to GPT-5.5, Claude Opus 4.7, and Gemini 3.
• Whether Meta releases the next major Llama generation (Llama 5 or equivalent) and whether it remains open-weights, given the strategic pivot toward closed-source.
• Integration of Scale AI's data infrastructure into Meta's training pipeline, including any commercial implications for Scale's existing customer base.
• Talent retention through the post-restructuring period, particularly in the MSL leadership cohort under Wang's authority.
• Yann LeCun's AMI Labs and whether its V-JEPA architecture produces the alternative-to-LLMs results LeCun has publicly predicted.
• AR and VR distribution strategy through Ray-Ban Meta smart glasses and Quest, including any first-party AI experiences that distinguish Meta's hardware from competitors.

Sources

• Fortune: Meta unveils Muse Spark, its first new AI model since hiring Alexandr Wang. April 2026 launch context.
• CNBC: Scale AI's Alexandr Wang confirms departure for Meta as part of $14.3 billion deal. 2025 restructuring announcement.
• The Decoder: LeCun exits Meta for his own startup. LeCun departure context.
• MIT Technology Review: Yann LeCun's new venture is a contrarian bet against large language models. AMI Labs context.
• Fast Company: LeCun says Meta "fudged a little bit" on Llama 4 benchmarks. Benchmark manipulation disclosure.
• Built In: Meta Superintelligence Labs explainer. MSL organizational reference.
• Wikipedia: Meta AI. Reference material.