ggml-org/llama.cpp b12179 was released on 2026-11-06. This repository is an LLM inference project implemented in C/C++. This release wires 3D convolution into the Hexagon path and adjusts tile-based execution and input unpacking computation. Improved handling of positive horizontal strides and better tile selection are also included. The source of these changes is the b12179 release: https://github.com/ggml-org/llama.cpp/releases/tag/b12179.
The core of this update is the wiring that lets the Hexagon path handle 3D convolution, together with adjustments to the related execution method. Tile-based execution and input unpacking computation were tuned together, and improved handling of positive horizontal strides and tile selection were explicitly named as improvement targets. So b12179 reads not as the addition of a single feature but as a change that handles convolution execution and tile handling on that path together. Since the repository is a C/C++ project for LLM inference, this entry focuses on changes in computation handling inside the inference path rather than on a general description.
The provided release summary presents no performance figures or benchmark results. Therefore it is not possible to judge actual speed improvements or effects on other execution paths from this change alone. What can be confirmed is limited to adjustments related to convolution wiring and tile-based execution on the Hexagon path, and performance effects beyond that are not supported by the given information.