There's been some success in creating translation layers that can convert between different LLM embeddings, and even between LLM and an image generation model.