Skip to main content
ZICQ
中 Log in / Sign up

Wiki Models & Products

Meta AI

Models & Products
Aliases： Meta AI FAIR Meta Superintelligence Labs ·2026-09-19

Meta AI

Meta AI is the AI division of Meta (Facebook's parent company), tracing back to FAIR (Facebook AI Research) founded in 2013 under Yann LeCun. In 2018 it merged with product-side AI; in 2025 it was reorganized under Zuckerberg's direction as Meta Superintelligence Labs, positioning "accelerating the path to superintelligence." FAIR was the first major company to push open-source LLMs — the Llama family gave open-source LLMs their first real chance to compete with GPT / Claude.

Historical milestones

  • 2013: FAIR founded; Yann LeCun as chief scientist
  • 2018: PyTorch 1.0 released, becomes the default academic framework
  • 2022-11: LLaMA 1, weights available only to research institutions
  • 2023-02: Llama 2 fully open-sourced (7B / 13B / 70B), sparks the open-source wave
  • 2024-04: Llama 3 (8B / 70B), first to approach closed-source flagships
  • 2024-07: Llama 3.1 (405B), the largest single open model
  • 2024-12: Llama 3.3 70B matches 405B at 6× smaller
  • 2025-04: Llama 4 fully on MoE; introduces Scout / Maverick / Behemoth
  • 2025: Meta Superintelligence Labs reorganized; poaches OpenAI / Anthropic researchers; "superintelligence" research framing

Key products

  • Llama family: see llama.md
  • PyTorch: de facto deep-learning framework
  • SAM / SAM 2 (Segment Anything): universal image segmentation
  • SeamlessM4T / Llama: multilingual translation and conversation
  • Make-A-Video / Emu Video: video generation
  • NLLB (No Language Left Behind): 200-language translation
  • AI Studio / Meta AI: consumer chatbot + assistant
  • AI Glasses (2024-): with Ray-Ban, integrating Llama multimodal

Influence

  • Llama made "open-source LLMs" go from "two years behind closed" to "on par or ahead"
  • PyTorch is the de facto standard in deep-learning research (90%+ academic papers)
  • "Large-scale open models" became Meta's brand signature
  • SAM / SAM 2 made universal segmentation possible
  • Llama multimodal in AR/VR glasses for on-device inference

Limitations

  • Post-Llama-4 release cadence slowed; community trust eroded
  • Consumer products (AI Studio etc.) trail ChatGPT / Claude in market penetration
  • The "superintelligence" narrative is criticized as Zuckerberg's personal obsession lacking a clear product path
  • Persistent controversy over training data copyright