Vibeleaderboard
Meta

ai lab

Also indexed as meta-ai · meta-llama · llama

Meta

Meta matters because few organizations combine a decade of highly visible basic research, frontier-scale compute, model weights that external developers can adapt, and consumer distribution measured in billions of users. The central uncertainty is strategic continuity: the FAIR and Llama record is established, while the balance between open-weight research and product-first Muse models remains unsettled after the latest reorganization.1,2,5,6

Official site ↗Models in the Atlas →

Profile

Overview

From Facebook AI Research to FAIR

Meta's modern AI research institution began in late 2013 as Facebook AI Research, commonly called FAIR. Mark Zuckerberg and Yann LeCun assembled the original team around long-horizon, openly published research, and the group later became Fundamental AI Research after Facebook changed its corporate name to Meta. FAIR's work ranged across computer vision, translation, speech, representation learning, robotics, and machine learning infrastructure rather than beginning as a product chatbot organization.1

Llama and the open-weight ecosystem

The Llama family moved Meta from a broad research publisher into a major foundation-model supplier. The first LLaMA release in February 2023 covered models from 7 billion to 65 billion parameters and made weights available to researchers. Later generations expanded access and commercial use, creating a large downstream ecosystem of fine-tunes, inference services, safety models, and deployment tools. Meta calls these models open source, but their custom licenses are not identical to standard open-source software licenses, so this dossier uses the more precise term open weight.2,1

Research beyond language models

FAIR also produced influential work outside language models. DINO showed that self-supervised vision transformers could learn strong image representations and emergent segmentation information without labels. Segment Anything paired a promptable segmentation model with the SA-1B dataset to support transfer across image-segmentation tasks. These projects illustrate a recurring Meta pattern: publish a general research artifact, release code or model assets, and let external researchers adapt them well beyond Meta's own products.3,4

The superintelligence reorganization

Meta reorganized its frontier effort again in 2025. The company invested in Scale AI, recruited its founder Alexandr Wang, and formed Meta Superintelligence Labs. Yann LeCun left at the end of 2025 to form a world-model startup. In 2026 the new organization released Muse Spark for Meta AI, followed by Muse Image and additional product integrations. Those releases establish that the rebuilt team is operating, but they do not yet settle whether Meta will resume broad open-weight frontier releases or prioritize models distributed inside its own applications.7,5,6,8

Company evidence

Epoch AI dataset ↗
Reported staff
3K
Jul 17, 2025[1]
Reported usage
1B monthly usersMeta AI assistant
Oct 29, 2025 · Confident[1]

Reported estimates, not audited figures. Confidence labels and source links are preserved from the dataset.

Reporting and context

Notable contributions

  1. 01Llama open-weight model distributionMeta released capable foundation-model weights from a company with frontier-scale resources and sustained the family across multiple generations, enabling a large independent fine-tuning and deployment ecosystem. This contribution concerns distribution and ecosystem scale, not a claim that Meta invented open-weight language models.2,1
  2. 02DINO self-supervised visual representationsThe DINO work demonstrated that self-distillation without labels could train vision transformers whose features supported strong classification and exposed semantic segmentation structure.3
  3. 03Segment AnythingMeta researchers released a promptable segmentation system and the SA-1B dataset, providing a reusable foundation-model approach to image segmentation across unfamiliar tasks.4
  4. 04Sustained open research from FAIRFAIR connected papers, code, models, and datasets across vision, translation, speech, and representation learning, preserving a broad research identity alongside Meta's product engineering organizations.1