A framework that decouples representational width from backbone width, offering the benefits of wider LLMs without the quadratic computational costs.