Apple Meta Muse Mac Studio

Apple and Meta spent years as rivals over privacy and ad targeting. On October 15 they announced a partnership to run Meta's 816-billion-parameter Muse Spark Open model locally on Mac Studios.

The numbers show who this is for. The setup links two to four Mac Studios with M5 Ultra chips and at least 1 TB of combined memory, reportedly for more than $40,000. That is a capital purchase, not a software subscription.

For enterprises, this changes the buying question. A frontier-class open-weight model on hardware you own means prompts and data never leave the building, and the cost is fixed instead of metered per token.

It also shows where the moat is moving. Meta released the weights under a license with limited commercial use. Apple supplies the silicon and updated its MLX framework. Neither is selling you tokens.

My read: the next AI procurement decision for regulated or IP-heavy teams will not be which API to call. It will be which workloads are worth owning the infrastructure for.

Which of your AI workloads would you move on-premises first if the model quality were good enough?