Meta FAIR
Last reviewed
Jun 3, 2026
Sources
23 citations
Review status
Source-backed
Revision
v1 · 1,565 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
23 citations
Review status
Source-backed
Revision
v1 · 1,565 words
Add missing citations, update stale details, or suggest a clearer explanation.
Meta FAIR (Fundamental AI Research), originally Facebook AI Research, is the long-standing fundamental research laboratory of Meta. It was founded in December 2013 with Yann LeCun as its first director, and it has built a reputation for open, publishable research and for releasing widely used software, models, and datasets, including the PyTorch deep learning framework and the early LLaMA language models. FAIR is distinct from Meta's product-oriented generative AI organization and, since mid-2025, sits inside the reorganized Meta Superintelligence Labs (MSL).[1][2][3]
Facebook announced the lab on December 9, 2013, naming LeCun, then a professor at New York University and one of the principal figures in deep learning, as its leader.[4][5] LeCun joined at the personal invitation of Mark Zuckerberg, and the lab was set up alongside Facebook's existing applied machine learning work as a dedicated home for long-horizon research.[1][6] LeCun became the first director of the research group, with the New York lab located about a block from NYU's campus in Manhattan; he stepped down from his directorship of NYU's Center for Data Science in early 2014.[5][7]
From the start the group was organized around a public, open-science model: Meta describes FAIR's mission as advancing the state of the art in artificial intelligence through open research for the benefit of all.[1][4] In practice this has meant publishing papers, releasing code and datasets, and open-sourcing models, which set FAIR apart from some industrial labs that kept more of their work proprietary.
Note that FAIR dates to 2013, not 1993; LeCun's earlier deep learning work, including convolutional networks for handwriting recognition at AT&T Bell Labs in the late 1980s and 1990s, predates the lab and was done elsewhere.[7]
FAIR's stated purpose is fundamental rather than product research: investigating the science of machine intelligence (computer vision, speech, natural language processing, reasoning, and self-supervised learning) and sharing the results openly. Many of its outputs became standard tools across the wider field. The lab has historically measured itself by published research and adopted open-source releases rather than by shipping consumer features, which is the role of Meta's separate applied and generative AI teams.[1][2]
FAIR began across three sites: Facebook's Menlo Park headquarters in California, its London office, and the new lab in New York City.[4][5] It opened a Paris lab in 2015, its first in Europe, and later added sites in Montreal, Tel Aviv, Seattle, and Pittsburgh.[8][9] As of 2025, Meta lists AI research locations in Menlo Park, New York City, London, Paris, Montreal, Tel Aviv, Seattle, and Pittsburgh.[2]
FAIR has produced a long line of widely cited models, datasets, and tools. The table below lists representative outputs.
| Year | Output | Area |
|---|---|---|
| 2015 | Faster R-CNN | Object detection |
| 2017 | PyTorch | Deep learning framework |
| 2017 | Mask R-CNN | Instance segmentation |
| 2018 | Detectron | Object detection platform |
| 2018 | fastMRI (with NYU) | Medical imaging dataset |
| 2019 | RoBERTa | Pretrained language model |
| 2019 | BART | Sequence-to-sequence pretraining |
| 2019 | wav2vec | Self-supervised speech |
| 2019 | Detectron2 | Object detection library |
| 2020 | wav2vec 2.0 | Self-supervised speech |
| 2021 | DINO | Self-supervised vision |
| 2022 | OPT (Open Pre-trained Transformer) | Open language model |
| 2022 | No Language Left Behind (NLLB) | Machine translation |
| 2023 | LLaMA | Open language model |
| 2023 | Segment Anything (SAM) | Promptable image segmentation |
| 2023 | DINOv2 | Self-supervised vision features |
| 2023 | SeamlessM4T | Multilingual speech and text translation |
PyTorch, released in 2017, became one of the most widely used machine learning frameworks and was later moved to an independent foundation.[2] In computer vision, FAIR's object-detection work (Faster R-CNN and Mask R-CNN) and its Detectron and Detectron2 code libraries were broadly adopted, as were the self-supervised DINO and DINOv2 image models and the Segment Anything Model (SAM) for promptable segmentation.[1][10] In language, RoBERTa and BART were influential pretrained models, and OPT was an early open large language model released with its training logbook.[2] In speech, the wav2vec family advanced self-supervised speech recognition, and the translation efforts No Language Left Behind and SeamlessM4T extended coverage to large numbers of languages.[1] The first LLaMA models were released in February 2023 as research artifacts; later, consumer-facing Llama releases came primarily from Meta's separate generative AI organization rather than from FAIR.[11]
LeCun led FAIR as director from 2013. In January 2018 Meta restructured its AI work: LeCun moved to the role of Chief AI Scientist, a more research-focused position, and Jérôme Pesenti joined as Vice President of AI, overseeing both the research lab and the applied machine learning group.[12][13] LeCun has said he wanted to step back from the administrative load of running the lab so he could focus on research.[13]
Joelle Pineau, a McGill University professor and reinforcement learning researcher, led FAIR for several years and became the Vice President of AI Research who ran the group day to day.[14][15] On April 1, 2025, Pineau announced she would leave Meta; her last day was May 30, 2025, after roughly eight years at the company.[14][15] In August 2025 she joined the AI company Cohere as Chief AI Officer.[16]
On May 8, 2025, Meta named Robert Fergus, a computer vision researcher who had spent about five years as a research director at Google DeepMind and who had been part of FAIR's early team, to lead the lab.[17][18]
FAIR's place inside Meta shifted several times. In June 2022 Meta decentralized parts of its AI organization and moved FAIR into the Reality Labs Research division.[19] Through 2023 and 2024, as Meta pushed into generative AI products, reporting noted that FAIR's profile inside the company had narrowed, with the high-profile Llama product releases coming from a separate generative AI team rather than from the research lab.[15]
In mid-2025 Meta carried out a larger reorganization. On June 30, 2025, it created Meta Superintelligence Labs (MSL), placing its AI efforts under Alexandr Wang, the former chief executive of the data company Scale AI, who joined as Chief AI Officer; the former GitHub chief executive Nat Friedman was named to lead AI products.[3][20] In August 2025 MSL was organized into four groups: a research unit known as TBD Lab focused on the Llama models and led by Wang; FAIR, the fundamental research lab, led by Fergus; a Products and Applied Research team led by Friedman; and an infrastructure team, MSL Infra, led by engineering vice president Aparna Ramani.[3] In October 2025 Meta cut roughly 600 roles across MSL, affecting FAIR, product AI, and infrastructure units while sparing the TBD Lab group.[3][21]
On November 19, 2025, LeCun confirmed he would leave Meta to start his own company focused on world-model architectures, systems intended to learn the structure and dynamics of the physical world rather than predict text; the venture has been reported under the name Advanced Machine Intelligence (AMI) Labs.[22][23]
As of early 2026, FAIR continues as the fundamental research group within Meta Superintelligence Labs under Fergus, separate from the product-focused TBD Lab and applied teams.[3]