Byte-Exact KV-Cache Grafting Increases Small Model Capability
July 14, 2026
KV-Cache Grafting allows frozen small language models to gain capabilities by restoring byte-exact key-value state artifacts from previously computed contexts. This method achieves 100% argmax agreement and zero KL divergence, moving a Gemma-4-12B from much lower scores to AIME 2025 performance levels without weight updates.
HOW THIS AFFECTS YOU
●
builderYou can dramatically improve model intelligence and reduce costs by grafting verified knowledge into new inference contexts.
●
founderThis presents a path to high-performing, low-cost AI products using smaller, frozen models.