Llama.cpp b11927 speeds up Gemma 4 12B 137% with multi-token drafter

The author demonstrates that enabling Google's multi-token prediction drafter for the Gemma 4 12B model in llama.cpp dramatically increases local inference speeds. By loading the 242-megabyte drafter file and applying four additional configuration flags, the author's RTX 4070 Ti Super setup achieved an average speed increase of 137 percent across JSON, coding, and prose workloads. Open article →