Meta AI

What it is

Meta AI is the artificial-intelligence research and product division of US company Meta Platforms Inc., headquartered in Menlo Park, California. It runs on two tracks: an open-weight model family (Llama) that any developer can download and deploy, and a consumer assistant embedded in Facebook, Instagram, WhatsApp, and Messenger, apps that together reach more than 3.5 billion daily active people. The division now houses two sub-units: FAIR (Fundamental AI Research), the original academic-style lab founded in 2013, and Meta Superintelligence Labs (MSL), created in June 2025 and oriented toward closed frontier development.

History

FAIR was founded in December 2013 when Meta recruited Yann LeCun, co-inventor of convolutional neural networks and 2018 Turing Award laureate, as chief AI scientist. The lab produced PyTorch, which became the dominant deep-learning framework in academic and industry research.

The Llama model series shifted Meta's standing in the AI industry. Llama 1 (February 2023) was a research release that leaked shortly after, seeding a wide open-source ecosystem. Llama 2 (July 2023) shipped under a formal commercial licence. Llama 3 (April 2024) trained on approximately 15 trillion tokens; its 405B variant (July 2024) was widely cited as the first open-weight model to reach frontier performance. Llama 4 Scout and Maverick (April 2025) introduced native multimodality via mixture-of-experts architectures: Scout (109B total parameters, 17B active) runs on a single Nvidia H100 with a 10-million-token context window; Maverick (400B total, 17B active, 128 experts) beat GPT-4o and Gemini 2.0 Flash on broad benchmarks at launch.

Mark Zuckerberg announced MSL via internal memo in June 2025, installing Alexandr Wang (former Scale AI CEO) and Nat Friedman (former GitHub CEO) as co-chief AI officers. LeCun, asked to report to Wang, declined and departed in November 2025. He launched AMI Labs in early 2026, raising US$1.03 billion at a US$3.5 billion valuation to pursue AI world models as an alternative to large language models.

Current state

As of mid-2026, Meta AI's consumer assistant counts roughly 600 million monthly active users, the largest publicly disclosed base of any AI assistant. India is its biggest single market. Meta Platforms' full-year 2025 revenue reached US$200.97 billion (up 22% year-on-year), with AI-driven advertising tools credited for a significant share of that growth. The company has guided 2026 capital expenditure to US$115-135 billion, most of it AI infrastructure including GPU clusters and data centers across the United States.

Muse Spark, released by MSL in April 2026, is Meta's first closed, proprietary frontier model. It is multimodal and benchmarked primarily on science, mathematics, and health reasoning. The release puts Meta squarely in the open-vs-closed frontier debate alongside OpenAI and Anthropic, and represents a deliberate break from the Llama open-weight tradition that built the division's developer base.

Relationships

Meta AI is wholly owned by Meta Platforms (see Meta Platforms), which also controls Facebook, Instagram, WhatsApp, and Messenger. The Llama weights are distributed through Hugging Face and the meta-llama GitHub repository, making Chinese labs, open-source fine-tuners, and enterprise vendors downstream dependants on Meta's model-release decisions. OpenAI and Anthropic are the primary closed frontier competitors. The acquisition of five Thinking Machines Lab founders in early 2026, led by Andrew Tulloch on a reported US$1.5 billion six-year package, marks the first major inter-lab talent competition triggered by MSL's formation.

What to watch

Whether the Muse family achieves sustained frontier parity with OpenAI and Anthropic on capability benchmarks over the next 12 months. Whether the Llama open-weight line continues under a new generation or is subordinated to the closed Muse programme. The US Federal Trade Commission's pending monopoly appeal could, if successful, constrain how Meta AI integrates with the social apps. Capital expenditure at US$115-135 billion in 2026 is among the largest single-company capital programmes in corporate history; power supply and GPU allocation from Nvidia are the principal execution risks.