Pattern: Multi-Token Prediction Merging Into Every Inference Stack as Standard Optimization

vLLM / llama.cpp·medium signal

MTP is no longer experimental — it's infrastructure. llama.cpp merged native support (May 16), vLLM already supports it via Gemma4MTPModel, and four major model families ship MTP heads trained from scratch (DeepSeek V3/V4, Qwen 3.6, Gemma 4). The pattern: MTP is following the trajectory of flash attention — from research novelty to expected default. For coding agent builders running local inference, MTP eliminates the need for separate draft models while delivering better acceptance rates (70–85% vs ~50%). Expect Ollama and other runtimes to add MTP support within weeks.