TL;DR: llama.cpp b10212 now loads Multi-Token Prediction (MTP) tensors only when actually used, trimming memory overhead for supported models.
Summary: Release b10212 of llama.cpp introduces a change (#26296) that loads MTP tensors only if they are really used, and skips loading MTP for remaining models that support it. This reduces memory consumption at load time, particularly for developers running MTP-capable models locally.
Why it matters: For AI builders deploying local LLMs, this is a memory optimization with no functional trade-off. Update your llama.cpp build and benchmark load-time memory if you use models with MTP support.
Source: github_releases