Manually turning it on is a bit troublesome, but the latency saved is really worth it.

CryptoWorld News: Draft models are beginning to phase out, and MTP (multi-token prediction) decoding capabilities are starting to be used in local inference front-end applications. The MTP approach involves adding several lightweight prediction heads to the main model, allowing the model to predict subsequent tokens in advance and verify them independently. The upstream model has already provided actions, and the DeepSeek-v3 technical report incorporates MTP into the training objectives, indicating that this module can be directly used to accelerate inference. Downstream inference frameworks and tools are also beginning to adapt, including llama.cpp, vllm, and LM Studio, with users required to download models that support MTP and enable it manually.