llama.cpp's b10212 release loads only needed MTP tensors

llama.cpp, the popular C/C++ inference engine for LLMs, rolled out b10212 on July 31. The release focuses on optimizing how MTP (multi-token prediction) tensors are handled. MTP lets models predict several future tokens at once, improving inference speed, but its associated weights were previously loaded unconditionally—even when the feature wasn't in use. Commit 82dbc4f changes this behavior to load MTP tensors only when the model actually requires them, and extends the skip logic to all remaining supported models.

For developers and self-hosters, this means faster cold starts and lower RAM consumption when running MTP-capable models without the MTP path enabled. The change is part of llama.cpp's ongoing effort to trim unnecessary I/O and memory footprint. The release also includes prebuilt binaries across platforms—macOS (Apple Silicon, Intel), Linux (x64, arm64, s390x, with Vulkan/ROCm/OpenVINO/SYCL backends), Windows (CPU, CUDA 12/13, Vulkan, OpenCL), and Android. While it's a small optimization rather than a headline feature, it's exactly the kind of practical win that keeps llama.cpp the go-to for efficient local inference.