Comment by DiabloD3

They weren't the first to do MTP like this, and arguably did it wrong: the MTP heads are kept in a separate file and have to be welded in by the inference engine.

Qwen 3.6 shipped with working MTP first, and had working MTP in llama.cpp first.