This seems to be a wrapper around llama.cpp with several tuned parameter settings. Most of the speedup is gained by enabling speculative decoding [1]. No amazing breakthroughs, just good tuning!
I have never understood how it is that "Ollama" manages to stay some sort of de-facto standard local hosting solution when all it does is wrap llama.cpp poorly. It's not as if llama.cpp is hard to use - it offers pre-compiled single file binaries! It implements every extant endpoint format! It has a web interface! What more could you ask for? Yet everything from agentic frameworks to vibe coded chat apps offer configurations for a variety of proprietary cloud providers, plus "Ollama". Never "llama.cpp". Why?