Meta AI

Meta AI is the artificial intelligence research and product organisation within Meta Platforms, the company that operates Facebook, Instagram, WhatsApp, Messenger, and Threads. It traces its origins to the founding of Facebook AI Research (FAIR) in 2013 under Yann LeCun, who remains Meta Chief AI Scientist. Meta AI today encompasses FAIR, the GenAI product organisation, and applied AI teams embedded across Meta product groups. It is among the most prolific producers of open-weight foundation models and open-source AI research in the industry.

Meta AI is best known publicly for the Llama family of open-weight large language models, the first generation of which was released in February 2023. Subsequent releases — Llama 2 (2023), Llama 3 (2024), Llama 3.1, 3.2, 3.3 (2024), and Llama 4 (2025) — have positioned the family as the most widely adopted open-weight base for fine-tuning across the industry. Meta also develops the Llama Guard safety models, the Code Llama family for code generation, and multimodal variants spanning vision, audio, and embodied agents.

Research output

FAIR has contributed seminal results to the field including PyTorch (released 2016 and now the dominant deep learning framework), the original DETR object-detection transformer, Segment Anything (SAM) and Segment Anything 2 (SAM 2) for image and video segmentation, ImageBind for cross-modal representation, DINO and DINOv2 for self-supervised vision, Wav2Vec and SeamlessM4T for speech, and the MMS multilingual speech recognition system. The lab publishes much of its work in peer-reviewed venues and releases model weights and code under permissive licences.

A second research thread under Yann LeCun pursues alternatives to standard generative training, including the Joint Embedding Predictive Architecture (JEPA) line of models and the broader programme on world models and autonomous machine intelligence. This work informs Meta longer-term agenda toward more general intelligence rather than larger language models alone.

Llama and the open ecosystem

Meta opens model weights under the Llama Community Licence, a permissive licence that allows commercial use up to defined scale limits. The decision to release weights — initially controversial after a leak of the first Llama checkpoint in March 2023 — has made Meta the most influential supplier of base models for the open ecosystem. Hugging Face statistics consistently show Llama derivatives as the most downloaded fine-tuned model family.

| Generation | Released | Notable variants | |---|---|---| | Llama 1 | Feb 2023 | 7B, 13B, 30B, 65B (research only) | | Llama 2 | Jul 2023 | 7B, 13B, 70B (commercial use) | | Llama 3 | Apr 2024 | 8B, 70B; long-context variants in 3.1, 3.2, 3.3 | | Llama 3.1 | Jul 2024 | 405B flagship, 128K context | | Llama 3.2 | Sep 2024 | Multimodal 11B and 90B vision, on-device 1B and 3B | | Llama 4 | 2025 | Mixture-of-experts flagship, native multimodality |

Consumer products

The Meta AI assistant, powered by the latest Llama variant, is integrated into Facebook, Instagram, WhatsApp, and Messenger, where users can chat with it directly or summon it inside group conversations. The assistant is available in many but not all jurisdictions; deployment is gated by local regulation and language coverage. In the EU, deployment has been adjusted to comply with the GDPR and the AI Act, and certain features remain limited.

AI Studio allows creators and businesses to build branded AI characters that interact with users in Meta apps. Ray-Ban Meta smart glasses, developed with EssilorLuxottica, integrate Meta AI for first-person photo and video understanding, real-time translation, and visual queries. Meta also operates a research-focused chat product at meta.ai and supplies AI features such as background generation, photo editing, and content recommendation embedded in its core social products.

Compute and infrastructure

Meta operates one of the largest AI compute footprints in the world. As of 2025 the company has reported clusters totalling hundreds of thousands of NVIDIA H100 and H200 GPUs across multiple data centres, alongside the in-house MTIA (Meta Training and Inference Accelerator) silicon for production recommendation workloads. The company has announced multi-year commitments exceeding USD 60 billion in capital expenditure to expand its AI compute base.

References

1. Touvron, H., et al. (2023). LLaMA: Open and Efficient Foundation Language Models. arXiv:2302.13971.
2. Meta AI. (2024). Introducing Llama 3.1: Our most capable models to date. ai.meta.com/blog.
3. Meta AI. (2025). Introducing Llama 4. ai.meta.com/blog.
4. LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. OpenReview position paper.