ggml-org/llama.cpp is a repository that performs LLM inference in C/C++. In the b11985 release, published on October 29, 2026, the way Metal's MXFP4 matrix-vector kernel and some operations are executed changed. This entry is based on the repository's b11985 release record.
According to the release summary, the MXFP4 matrix-vector kernel now unpacks values using an FP16 bit technique, so it no longer uses threadgroup memory. Unary operations, binary operations, and GLU rows were changed to be executed split into larger threadgroups. The first change concerns how memory is used, and the second concerns the size at which operations are split. The scope of this update is how Metal handles this kernel and these operations. The summary does not spell out the specific processing steps or any user-visible functional changes.
The provided release summary contains no figures comparing processing time, throughput, or memory usage. Therefore, it is not possible to judge from this information alone how the change that stops using threadgroup memory, or the approach of splitting into larger threadgroups, affects performance. The summary also covers Metal changes and does not state whether other backends were changed. What can be confirmed in b11985 is a change in how operations are executed on the Metal path, and the size of its effect cannot be determined from the given facts.