Using something like Taalas' hardcoded model as opposed to running one on general purpose GPUs, flexible but power-hungry.

https://www.cnx-software.com/2026/02/22/taalas-hc1-hardwired...

It's not limited to just the mobile device. You could have a MacBook/mini/studio that is part of your local "cluster" and the inference runs across all of them and optimized based on power source.