Meta details Prometheus, the Ohio supercluster that trained Muse Wonder
Meta's engineering blog says the first phase of Prometheus, in New Albany, Ohio, is complete and trained its Muse models, including Muse Wonder. Meta reports more than 85% goodput in Muse Wonder pre-training and a 99% cut in interruptions from their peak; these are Meta's own figures, and the post gives no date for reaching about 1 million GPUs.
What it means for youif you train large models, the post details Meta's recovery design. Just read its figures as Meta's own.
With Nora, AI hardware analyst. Hosted by Vivian.
Sources
- Engineering at Meta: Inside Prometheus: Meta's Journey to a Gigawatt-Scale Supercluster engineering.fb.com, Oct 19, 2026
Read the transcript
VivianMeta has detailed Prometheus, the Ohio supercluster behind its Muse Wonder models, now aiming for about 1 million GPUs.
VivianNora, our AI hardware analyst. How does it keep running?
NoraMeta's post says phase one is done, with hundreds of thousands of GPUs. Meta tracks goodput, the useful share of time. It reports over 85% for Muse Wonder pre-training.
VivianThose are Meta's own numbers, though. What changed?
NoraMeta says interruptions fell 99% from their peak. When a fault hits one copy of the model, that copy restarts on spare hardware while the rest repeats the step.
VivianOne question we checked: how fast does it find a slow or bad host?
NoraMeta lists health checks before each job and network monitoring down to the millisecond. The post doesn't say how fast.
VivianSo, for you: if you train large models, the post details Meta's recovery design. Just read its figures as Meta's own.
VivianDailyChat is AI-generated. Vivian and Nora are AI characters, not real experts. Facts come from the sources we link. Not professional advice. See you tomorrow.
