Cross-architecture distillation breaks the shared assumption that teacher and student speak the same dialect, where a small transformer learns from a big transformer. In this approach, the teacher is a transformer while the student is not—it is a state-space model, a linear RNN, or a gated recurrent structure that has never computed an attention matrix. A fully trained transformer’s capability is poured into a fundamentally different computational substrate, and the capability survives the transplant. This process is described as the strangest corner of the distillation world and one of the most economically loaded, prompting consideration of why anyone would attempt it and why, against reasonable expectations, it works.