llama.cpp b10228 Released with Support for DeepSeek-V4 MTP and DSpark
The b10228 release of llama.cpp introduces support for DeepSeek-V4's Multi-Token Prediction (MTP) and DSpark. This update integrates these features directly into the local LLM inference engine. This release enables local deployment of DeepSeek-V4 models with significantly accelerated inference speeds. By leveraging MTP and DSpark, users can run highly efficient speculative decoding without needing a separate draft model. The update is tracked under pull request #25784 and is available across multiple platforms, though KleidiAI support for macOS Apple Silicon remains disabled in this build.
## BACKGROUND
Multi-Token Prediction (MTP) is a speculative decoding method where a language model natively predicts multiple future tokens at once to speed up generation. DSpark is an inference optimization technology developed by DeepSeek that significantly accelerates model execution. llama.cpp is a popular open-source C/C++ library designed for running large language models locally with high efficiency.