Linear Mapper Transfers KV Cache Between LLMs at Up to 25x Faster Than Re-Prefill

Researchers have proposed a method to transfer key-value (KV) caches between different-sized large language models in the same family without re-running the prefill process from scratch. The approach uses a closed-form ridge regression mapper trained on just 500 calibration sequences to project KV caches from a source model into the target model's space. Tests on model pairs such as Qwen3 14B to 32B showed speedups of 2.7 to 25 times over standard re-prefill, while retaining 73 to 98 percent accuracy across most pairs. The pipeline involves selecting the most predictive source layers, stripping rotary position embeddings to make the mapper position-independent, and then applying a ridge regression fit per attention head. The technique is particularly valuable in production systems that use model cascading, mid-conversation switching, or dynamic routing, where repeated prefill costs can add up significantly.