Long-context serving relies on KV caches to avoid recomputing past tokens. But that efficiency disappears when the model changes: the cache no longer applies, and the receiving model must prefill the entire context again. A new arXiv paper, CacheLink, frames this as a problem of cross-model context handoff and asks whether the prefill step can be eliminated entirely.

The paper introduces the concept of functional KV transport, where cached state is moved or transformed so a different model can continue from where the previous one left off. The authors call this zero-prefill cross-model context handoff, and CacheLink is presented as a mechanism to enable it. The abstract does not give implementation details, but the motivation is clear: model switching should not force a costly restart of the context.

If the approach works, it could make multi-model pipelines and speculative or cascading inference far cheaper, since each model change would no longer require a full re-read of the conversation. The paper is a research announcement, so the claims are preliminary and not yet validated by the experiments described in the full text.