Meta AI

Meta AI

Overview

Meta AI evolved from Facebook AI Research (FAIR), founded in 2013 as the company's dedicated artificial intelligence research division. Originally built on a foundation of open-source philosophy and fundamental ML research, Meta AI has undergone a strategic pivot toward frontier large language models. Under the leadership of [[Yann LeCun]], Chief AI Scientist, the lab established itself as a major producer of state-of-the-art models, most notably the Llama family of large language models. This trajectory reflects Meta's broader shift from a social media company to an AI-first organization, culminating in the formation of Meta Superintelligence Labs in 2025 and a dramatic increase in AI capital expenditure to compete aggressively in frontier model development.

In April 2026, Meta executed a sharp strategic reorientation. The company unveiled Muse Spark — the first flagship model from Meta Superintelligence Labs under Chief AI Officer Alexandr Wang — and explicitly positioned the new "Muse" series as a successor to Llama, breaking from the previous open-source-by-default playbook. Muse Spark is closed-source, voice/text/image input with text output, and powers the standalone Meta AI app and desktop website (with Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta glasses rollouts following). The 2026 fiscal year marks an inflection point for Meta AI, with announced AI capex reaching $115–135 billion—more than double 2025's $72.2 billion.

Key Details

  • Founded: 2013 (as Facebook AI Research / FAIR)
  • Chief AI Officer: Alexandr Wang (Scale AI co-founder)
  • Chief AI Scientist: [[Yann LeCun]]
  • Headquarters: Menlo Park, CA
  • Parent Company: Meta Platforms Inc.
  • 2026 AI Capex: $115–135B (more than double 2025's $72.2B)
  • Website: https://ai.meta.com

Current Models

  • Muse Spark: First flagship from Meta Superintelligence Labs; multimodal input (voice + text + image), text output; closed-source; competitive with frontier proprietary models at lower compute cost than mid-size Llama 4 (April 8, 2026)
  • [[Llama 4 Maverick]]: 17B active parameters / 400B total; 1M context window; beats GPT-4o; released April 5, 2025
  • [[Llama 4 Scout]]: 10M context window on single H100; 17B active / 109B total; 16-expert mixture-of-experts; released April 5, 2025
  • [[Llama 4 Behemoth]]: Flagship unreleased model; 288B active / 2T total parameters; outperforms GPT-4.5 on STEM benchmarks
  • Mango: Multimodal image/video generation model (H1 2026 planned)

Key People

  • Alexandr Wang: Chief AI Officer; head of Meta Superintelligence Labs; co-founder of Scale AI; recruited via Meta's $14.3B investment for a 49% stake in Scale; led Muse Spark development over 9 months
  • [[Yann LeCun]]: Chief AI Scientist; deep learning pioneer and Turing Award winner (2018, shared); strong advocate for open-source LLMs and skeptic of certain paths to AGI; in May 2026 publicly called CEO-driven AI job-loss narratives "extremely destructive" and pushed back on Dario Amodei's predictions about AI replacing entry-level white-collar work
  • Various FAIR researchers: Large team of ML researchers and engineers across Menlo Park and distributed offices

Recent Developments

  • Muse Spark Launch (April 8, 2026): First model from Meta Superintelligence Labs under Alexandr Wang — built over 9 months and originally code-named "Avocado." Closed-source (a deliberate departure from Llama's open-weight playbook), with a planned open-source variant to follow. Powers Meta AI assistant in the standalone app and desktop site, with Facebook/Instagram/WhatsApp/Messenger/Ray-Ban Meta rollout coming. Meta describes Muse Spark as competitive with frontier proprietary models on multimodal perception, reasoning, health, and agentic benchmarks — at a fraction of the compute cost of Meta's older mid-size Llama 4 variant.
  • AI Capex Confirmed at $115–135B: 2026 capital expenditure on AI infrastructure confirmed at $115–135 billion, more than double 2025's $72.2B, paired with the Muse Spark launch.
  • Strategic Open-Source Shift: The Muse Spark closed-source default marks Meta's clearest pivot away from the "Linux of AI" Llama positioning, though Meta has stated it will release an open-source variant of Muse Spark in the future.
  • Llama 5 Reporting (April 2026): Earlier reporting indicated a Llama 5 release in early April with 600B+ parameters and 5M-token context. The subsequent Muse Spark launch is widely interpreted as either a re-naming or a new flagship effectively superseding the Llama line going forward.
  • Scale AI Stake (2025): Meta acquired a 49% stake in Scale AI for $14.3B, recruiting Alexandr Wang to lead Superintelligence Labs.
  • NVIDIA Partnership Expansion: Estimated $35–67B in chip commitments.

Why They Matter

Meta AI represents the largest frontier AI effort of a major social media platform, with unparalleled computational resources and capital commitments. The 2026 strategic pivot — Muse Spark as a closed-source flagship under Alexandr Wang — represents one of the most consequential repositionings in the AI industry: Meta has effectively conceded that open-source-by-default is not a winning strategy at the frontier, while preserving an open-source release path to retain ecosystem credibility. Meta's 2.7 billion+ daily users across Facebook, Instagram, WhatsApp, Messenger, and Ray-Ban Meta represent the largest AI distribution surface in the world, and Muse Spark gives the company a frontier-grade model to put behind that surface for the first time. With $115–135B in 2026 AI capex and Wang's Superintelligence Labs running in parallel to LeCun's FAIR, Meta is now structured to compete simultaneously with OpenAI and Anthropic on capability, with Apple on consumer assistants, and with Google DeepMind on multimodal foundation models.

Last Updated

May 7, 2026

→ Back to Companies