This technical benchmark evaluates the performance gains of Multi-Token Prediction (MTP) using Gemma-4-12B-it on consumer-grade hardware (NVIDIA RTX 3060 12GB) via llama.cpp. The implementation utilizes an external 4-bit quantized Assistant model as a drafter to speculate up to four tokens ahead. Results indicate that while MTP slightly reduces token ingestion speeds to 70–73% of baseline, it approximately doubles token generation speeds. Specifically, in a four-turn coding task, MTP reduced total execution time from 293,705 ms to 173,150 ms, representing a 40% efficiency increase. Generation speeds improved from a baseline of ~34 tps to ~62 tps. The study observed a high average token acceptance rate of 81.06% during structured tasks like code generation, likely due to the predictable grammar of programming languages. Furthermore, the Gemma tokenizer's large 256k vocabulary allows it to treat Japanese text as whole words, facilitating high-speed MTP performance even in non-English contexts. The experiment confirms that MTP provides significant throughput benefits for agentic workflows and code generation by effectively utilizing speculative execution principles within the inference engine.