An XDA writer reports that pairing Gemma 4 12B with its matching MTP drafter in llama.cpp increased average generation speed from 75.3 to 178.7 tokens per second. The personal test compared JSON, code and prose tasks on an RTX 4070 Ti Super with otherwise matching quantization and context settings. A small model proposes multiple tokens for the larger model to verify, requiring compatible files and runtime support. This is a practical use of an already-released feature rather than a new model launch. Gains vary with hardware, workload and model architecture.