Preprint reports faster context transfer between different AI model families

Researchers propose HeteroFold, a method that lets different AI model families reuse previously processed context without having the receiving model process the text again. The preprint reports that transferring a 32K context from Llama-3.1-8B to Ministral-3-14B was 10.7 times faster than native context processing and 1.18-1.47 times faster than two existing transfer methods. HeteroFold aligns model structures, translates the sender's stored attention data into the receiver's representation, and calibrates it while keeping both models frozen. It matched text-based communication on a multi-agent benchmark. The reported speed comparison applies to the specified model pair and context length.