A Reddit user shared their experience with the Gemma 4 31B QAT (Quantization-Aware Training) model. Compared to traditional GGUF quants like Q6_K_L, the QAT version delivers noticeable quality improvements in roleplay and long-context tasks. Additionally, combining the QAT model with Multi-Token Prediction (MTP) yielded massive speedups, boosting generation speeds from ~20 t/s to up to 50 t/s.
A popular Reddit thread addresses user confusion over running Gemma 4 31B locally. It distinguishes between MTP (Multi-Token Prediction for inference speedup) and QAT (Quantization-Aware Training for preserving 4-bit quality). It also confirms that llama.cpp's new MTP support requires updated GGUF files and a secondary draft model file for acceleration.