Google strips MTP heads from public Gemma 4 weights

Google's decision to decouple MTP heads from public weights is a strategic move that prioritizes ecosystem control over true open-source parity. MTP enables 1.5x-2.0x faster inference through built-in speculative decoding, a major advantage for on-device AI. Stripping these heads from Hugging Face weights ensures that developers seeking maximum performance must use Google's LiteRT framework. While Google cites compatibility as the reason, the move creates a two-tier system that hinders third-party optimization in tools like llama.cpp. The community is already working on reverse-engineering the LiteRT models to stitch MTP support back into standard formats.