ggml-org/llama.cpp is a C/C++-based LLM inference project. In b12125, released on November 5, 2026, the multi-column GEMV path for f32 and q5_k is enabled by default on the Adreno X2 series. The change is recorded in the b12125 release information on GitHub. Therefore, the core of this release is that the default selection within the OpenCL processing path has changed. The main change of the new version is not a general feature addition, but the switch of that computation path to the default for a specific GPU series.
The change targets the OpenCL processing path of the Adreno X2 series, and what is enabled by default is the multi-column GEMV path for f32 and q5_k. f32 and q5_k are the names of the two formats given in the release description, and GEMV is the name of the operation of the path being changed. What the release information confirms is the fact that this path has been enabled by default and the GPU series listed as the target. Therefore, when reading the contents of b12125, one should focus on the scope in which the new default selection applies.
The release description contains no performance measurements for this change, nor any comparison results with the existing path. How fast the new default path is, and under which conditions differences arise, are not given either. Whether the same switch applies to other GPU series or to processing paths other than OpenCL cannot be confirmed from this information alone. Therefore, this change should be understood as limited to an update in which the default OpenCL path of the Adreno X2 series has changed, and there is no basis for regarding it as a release that has demonstrated a performance gain.