SwiftKV reduces prefill compute by up to 50% by combining model rewiring and knowledge-preserving self-distillation.