HomeFull speed ahead. Keep track of advancements in every frontier

TheTechno Optimist.

Prometheus and HyperionMeta—Prometheus and HyperionMeta—Muse modelsMeta—Meta—Muse modelsMeta—Muse modelsMeta—Muse modelsMeta—

Meta

Llama open models, Muse, AI data centers

USpublic~74,067 staffest. 2004ai.meta.com (opens ai.meta.com)Checked 25 Sep

The brief

Analyst view
  • Llama was downloaded over a billion times, but Llama 4 (April 2025) was seen as behind the leaders.
  • Paid about $14.3B for 49% of Scale AI and hired its CEO Alexandr Wang to lead its new AI lab (June 2025).
  • Muse Spark (April 2026) is closed, a shift away from open weights.

Technical approach

As reported
Open weights

Giving models away

Llama's weights are published, so developers can run and adapt them. Thousands of other models are built on Llama.

Architecture

Mixture of experts

Llama 1 to 3 used the whole model for every word. Llama 4 switches on only some parts, and handles images natively.

Compute

Its own superclusters

Meta trains on its own GPU clusters, from 24,000-GPU clusters in 2024 to gigawatt campuses in Ohio and Louisiana.

FAIR

World models

Meta's FAIR lab builds models like V-JEPA 2 that learn how the world works from video, used for robot planning.

Programs

Meta, three programs

Updated 17d ago
LlamaMeta's open-weight AI models, free to download and allowed for most commercial use. Llama 4 added mixture of experts and built-in image understanding.ScalingUpdated 5 Apr 2025View program →Muse modelsModels from Meta Superintelligence Labs. Muse Spark (code-named Avocado) focuses on reasoning and coding and powers Meta AI. Meta hopes to open future versions.DeployUpdated 17d agoView program →Prometheus and HyperionMeta's giant AI training sites, announced in July 2025: Prometheus in Ohio and Hyperion in Louisiana, planned to reach up to 5 GW over several years.DeployNextPrometheus online · 2026Updated 13 Jul 2026View program →
Research · Proto · Pilot · Deploy · Scaling

Research & demos

4 items
  • Jun 2025
    paperV-JEPA 2

    Learned how the world moves from video, then planned robot actions without extra training.

  • Jul 2024
    paperThe Llama 3 Herd of ModelsLlama

    Explained how a 405B dense model was trained on about 15 trillion tokens using 16,000 H100 chips.

  • Apr 2023
    paperSegment Anything

    A model that cuts out any object in an image on request, plus over 1 billion example masks.

  • Feb 2023
    paperLLaMA: Open and Efficient Foundation Language ModelsLlama

    Showed smaller models trained on more data can match bigger ones. It started the open-weight wave.

Open technical problems

Open · Active · Partial
  • Catching the frontierOpen

    Llama 4 trailed the best closed models. A special version used on a public leaderboard differed from the release, hurting trust.

  • Open versus closedOpen

    Meta must weigh the benefits of open weights against competition and safety. Muse has not been opened.

  • Return on spendingOpen

    AI infrastructure spending runs well above $60B a year. Payback comes mostly indirectly, through ads.

Peers

Open-weight models · 4

Company & funding

Staff
~74,067as of 2024
HQ
US
Founded
2004

Partners & customers

3 partners · 2 suppliers
  • Partner
    Scale AI

    Meta took a 49% non-voting stake for about $14.3B in June 2025.

  • Supplier
    NVIDIA

    Main GPU supplier for training clusters.

  • Supplier
    CoreWeave

    About $14.2B cloud capacity agreement (Sep 2025).

  • Partner
    Blue Owl Capital

    Joint venture financing the Hyperion data center (Oct 2025).

  • Partner
    Microsoft

    Launch partner for Llama 2 on Azure (2023).

People

Key people
  • Mark ZuckerbergFounder, Chairman and CEO
  • Alexandr WangChief AI Officer
Primer