LatentPort: Beyond KV Cache - Cross-Model Transfer of Recurrent Memory in Hybrid Language Models: A 4B-to-9B Hybrid-State Handoff Without Target Prefix Replay

LatentPort. Beyond KV Cache - Cross-Model Transfer of Recurrent Memory in Hybrid Language Models: A 4B-to-9B Hybrid-State Handoff Without Target Prefix Replay. By Simon P. Villani. Published on the seventh of September. Its code repository is worked out: about ninety source files, with a readme and a licence.

Can one language model hand its live memory to another, without the receiver rereading the context? The author shows that it can, for one specific pair: an architecture-matched Qwen3.5 sibling pair, from four billion parameters to nine billion. To the author's knowledge, this is the first demonstrated cross-model handoff of persistent recurrent inference state between differently sized hybrid language models without target prefix replay. The larger model picks up where the smaller one left off, without processing the history itself. The title's phrase for this is a hybrid-state handoff: what is passed across is not only the attention cache, but the persistent recurrent state that hybrid models also keep.

Translating the attention KV cache alone leaves a large gap. Adding the persistent-state package from Gated DeltaNet, or GDN, lowers teacher-forced negative log-likelihood, the average next-token log-loss, by zero point seven four seven nats per token. The ninety-five percent paired bootstrap interval runs from zero point six nine two one to zero point eight zero four seven, and every one of the sixty-four PG19 documents improves.

Reusing the recurrent and convolution state directly works better than the learned GDN maps tested, which is consistent with the persistent-state coordinates being partly compatible between the two models. A fresh component factorial selects translated KV together with direct recurrent and convolution state.

A further correction with four hundred and thirty-four thousand one hundred and seventy-six parameters improves on that base, measured on sixty-four fresh web documents. Continuation loss sits zero point zero seven six nats per token above native nine billion. Jensen-Shannon divergence is zero point zero two two. Native context recovery is zero point nine one eight. And the corrected nine billion model significantly beats simply continuing inference on the four billion model, while processing zero historical prefix tokens.

The author sets the limits carefully. The evidence covers one direction, one geometry-matched Base-model pair, and continuation up to four thousand tokens under teacher forcing. The near-native gate failed. The sixteen thousand token branch was not run. Equivalence under free generation, and a general interface for state, both remain unproven.

Within those limits, the handoff works: live memory crossed from a smaller model to a larger one, and the larger model was better off with it than without it.