GoKawiilMeta details Prometheus supercluster used to train Muse foundational models
Meta says the first phase of its gigawatt-scale Prometheus supercluster in New Albany, Ohio is complete and was used to train its Muse family of foundational models, including the new Muse Wonder models. The company plans to scale the cluster toward roughly one million GPUs and says it uses closed-loop liquid cooling, having achieved over 85% goodput in pre-training and over 90% in reinforcement learning runs for Muse Wonder.
GoKawiil's interpretation of the reporting above, not reported fact.
The disclosure signals Meta's push to compete in frontier AI infrastructure at a scale rivaling other major labs, suggesting heavy capital commitment to training capacity. The reported reliability metrics, like goodput and mean time between interruptions, indicate engineering focus on squeezing more usable compute out of massive GPU fleets, which could translate into faster model iteration.
- Prometheus's first phase is complete and already trained Meta's Muse Wonder models
- Meta aims to scale the supercluster toward about one million GPUs
- Reliability improvements raised mean time between interruptions from ~47K to over 54K GPU hours
Source: engineering.fb.com — Kevin Lee, 2026-10-19
Published there as: “Meta's GB{2,3}00 Cluster”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.