Cache-to-Cache lets AI models swap internal memory directly, boosting inference speed by 150%

Researchers at Tsinghua University published a paper on Cache-to-Cache (C2C), a technique that lets separate AI models exchange internal memory without generating text. Accepted at ICLR 2026 with open-source code, it delivers 100–150% faster collaborative inference and up to 14.2% accuracy gains.

• C2C passes one model's cache into another's memory via a learned Fuser module
• Selective gating decides which layers absorb incoming information
• Collaborative tasks run 100–150% faster, accuracy improves 3.1–5.4%
• Works only with open-weight models due to internal cache access