Researchers from Tsinghua University have introduced the Cache-to-Cache (C2C) method, which allows AI models to communicate directly without generating text. The technology enables one model to transfer the contents of its operational memory, or cache, directly to another model, bypassing the need for textual conversion.
The process is managed by a software component called Fuser, which adapts data between models with different internal structures. An intelligent filter within the C2C framework selects which data fragments to integrate immediately, while other model layers continue autonomous processing. This approach eliminates the computational overhead associated with human-like text generation.
According to the researchers, implementing C2C increases the speed of AI models performing joint tasks by 100–150 percent. Furthermore, the method improves accuracy by 14.2 percent compared to autonomous operation and by 3.1–5.4 percent compared to traditional text-based exchange.
Currently, the C2C method is only compatible with open-weight AI models, as it requires direct access to the cache and layer architecture. The source code for the development has been made available to the public.
