Meta Muse: the model family trained on the Prometheus supercluster

Meta has released the Muse family of foundational models, and this announcement is about the infrastructure behind them. The models were trained on the first phase of Prometheus, Meta's gigawatt-scale supercluster in New Albany, Ohio. Meta describes Muse Wonder as its new frontier models for personal and professional agentic tasks.

Meta says it reached more than 85% goodput on pre-training runs and more than 90% on reinforcement learning runs of Muse Wonder. Mean time between interruptions was about 47k GPU hours during those runs. Firmware fixes have since raised that to more than 54K GPU hours, according to Meta.

Prometheus currently runs two NVIDIA platforms: Catalina, built on GB200 NVL72, and Clemente, built on GB300 NVL72. A third platform based on Vera Rubin NVL72 is arriving soon. Meta says the capacity lets it pursue larger models and longer context windows than were previously practical.

Meta is still scaling Prometheus with its next frontier models from Meta Superintelligence Labs in mind. At full scale it is meant to reach roughly 1 million GPUs. Muse Wonder Full is the one earlier model listed from the lab.