speculative-decoding

Speculative Decoding Explained: 2x Faster Local LLMs with Ollama and llama.cpp 2026

Speculative decoding doubles local LLM inference speed with zero quality loss. How it works, how to enable it in Ollama and llama.cpp today, and which model pairs give the best speedup.