Meta and Apple just put an 816B parameter model on your desk, and the math is completely broken.
Muse Spark Open is technically impressive. 816B parameters, 42B active, 1M context, open weights, MLX optimized. But running it locally takes roughly 1TB of unified memory. In hardware, that means two 512GB M5 Ultra Mac Studios or four 256GB ones. Call it $38,000.
A frontier API subscription runs $200 to $500 a month. For $38,000, you can subscribe for six years and get the newest model every year. Open weights lag the frontier by 6 to 9 months anyway. The hardware buys you yesterday’s capability at tomorrow’s price.
Data sovereignty does not rescue the pitch. It is a contract problem now, and zero retention agreements solve it for a fraction of the cost. To justify owning hardware, you need 24/7 saturation. But the same compliance rules that force data onto a device also demand human approval, locked screens, and working hours. Your private cluster sleeps 16 hours a day. Nobody serious scales inference on macOS anyway. FileVault alone ends that experiment.
Uncensored, fully controlled weights are the only real upside that survives. Valuable for researchers and red teams. Not a $38,000 argument for the rest of us.
This is a gift-wrapped AI PC narrative for Apple, an ecosystem play for Meta, and a beautiful toy for everyone else.
Frontier-class open models on unified memory will get smaller and cheaper. The minute a distilled version runs on one machine, the math changes. Until then, subscribe. Do not buy.