LatentPort Enables Cross-Model Memory Transfer Between Qwen 4B and 9B
September 6, 2026
LatentPort enables the transfer of persistent recurrent inference state between different sized Qwen3.5 models without requiring prefix replay. Utilizing Gated DeltaNet (GDN) lowers teacher-forced NLL by 0.747 nats/token during 4B-to-9B handoffs.
HOW THIS AFFECTS YOU
●
builderYou can optimize inference by handing off live memory from smaller models to larger ones without the latency of re-reading context.
●
researcherThis demonstrates a new method for cross-architecture recurrent state transfer using GDN packages.