NVIDIA published a case of running Meta's Muse Spark on GB300, but any link to the same week's research papers is not confirmed.
An NVIDIA technical blog described how to run Meta's Muse Spark on GB300 NVL72 to scale agentic coding workflows. Four arXiv papers on mixture-of-experts and self-correction also appeared the day before.
Optimizing Meta's model for the latest rack-scale platform may signal that inference-infrastructure choices are moving into production. However, the source is NVIDIA, Meta's own adoption scale is not confirmed, and any direct tie to the papers is unverified.
Watch for an official Meta statement, deployment scale, and measured GB300 performance and cost figures. Also track whether MoE research moves into real implementations.