Apparently it is the same as the DeepseekV3 architecture and already supported by llama.cpp once the new name is added. Here's the PR: https://github.com/ggml-org/llama.cpp/pull/18936