Update on Gemma 4 having MTP: Reverse engineering effort

In a significant development for the open-source AI community, a developer has initiated a reverse-engineering project targeting Google's upcoming Gemma 4 model. The individual, operating under the username Electrical-Monitor27, successfully extracted the model's weights and discovered it contains MTP (Multi-Task Planning) capabilities—a feature that allows a single model to handle multiple, complex planning tasks. The extracted files are in the form of compiled TFLite (TensorFlow Lite) graphs, which are quantized in INT8 format, suggesting Google may have used Quantization-Aware Training (QAT). This quantization could make the model more efficient but also presents a challenge for reconstruction.

The developer has created a public repository on HuggingFace containing all extracted files, including a JSON representation of the Graphdef, which details the model's architecture. The core challenge now is converting these low-level TFLite files back into a high-level, usable PyTorch `nn.Module`. Electrical-Monitor27 is specifically calling for assistance from developers proficient in C++ to lead this conversion effort. The project leverages clues from previous community efforts, like the Gemini Nano extraction, and suggests using tools like Google's own AI Edge Model Explorer to aid the reverse-engineering process. The end goal is to salvage and port Gemma 4's MTP capabilities into an open, functional framework.