Rather than scaling down a model to fit multiple small devices, I like the idea of splitting a bigger modelʼs layers across that same set of small devices.

Something like https://github.com/orailnoor/cross-platform-llm-client/relea...