Technology ┬╖ News brief

Meta releases Muse Spark Open with 816B total parameters

Meta's own research announcement describes Muse Spark Open as an 816-billion-parameter mixture-of-experts model with about 42 billion active parameters per token and a one-million-token context.

Meta Platforms campus, Menlo Park. Archived/contextual photo; not the newly reported event.
File photo (not the reported event): Meta Platforms campus, Menlo Park. Credit: Wikimedia Commons; archival headquarters photograph. Image source and licence

What happened

Meta's own research announcement describes Muse Spark Open as an 816-billion-parameter mixture-of-experts model with about 42 billion active parameters per token and a one-million-token context.

Context and what is known

It is an open-weight release, not unrestricted public-domain software. Developers should read the model license and hardware requirements before deploying it.

What the announcement means

Meta's release refers to accessible model weights, not a guarantee that the system is effortless or inexpensive to run. The announced 816-billion total parameters differ from the smaller number active per token; memory and hardware needs still matter. Developers should also read the model's actual licence and safety documentation before treating open-weight availability as the same thing as an unrestricted open-source service.

Reporting source

Meta AI Research тАФ original reporting

тЖР Homeрд╣рд┐рдВрджреА рдореЗрдВ рдкрдврд╝реЗрдВ тЖТ