I tested MTP on vLLM and llama.cpp for Gemma 4 & Qwen 3.6 — 3.34x faster inference, here are my findings RTX 6000 PRO.

In three linesMTP (Multi-Token Prediction) benchmark on Gemma 4 31B and Qwen 3.6 27B using vLLM and llama.cpp. Result: 3.34x speedup (132.52 vs 39.69 tok/s). vLLM outperforms llama.cpp on Gemma 4; llama.cpp solid on Qwen. No confirmed quality degradation, VRAM overhead negligible.

Summary generated by Claude — human-verified