ggml-org/llama.cpp b11866 is a new build of llama.cpp, a C/C++-based LLM inference project. Released on October 24, 2026, this version removed the maximum-value rescaling from the mixture-of-experts matrix operation in the Metal backend. The change is recorded in the b11866 release information. In the computation graph, it was replaced with a per-expert, token-level processing method. As a result, the temporary buffer and the two-stage reduction kernel were also removed.
The scope of this change is the processing path of the mixture-of-experts matrix operation. The previous maximum-value rescaling is gone, and the computation graph now processes per expert at token granularity. The release information explicitly lists the removal of the temporary buffer and of the reduction kernel that had proceeded in two stages. Therefore, what can be confirmed is a change to the operation path and its components; the given information contains no statement that new inference functionality was added. The project itself only provides a description that it performs LLM inference in C/C++.
The provided release information lists no performance figures and no model used for validation. Therefore, it is not possible to judge how much processing speed or memory usage changes with this change. What can be said from the published facts is limited to the fact that the maximum-value rescaling, the temporary buffer, and the two-stage reduction kernel of that operation have been removed from the Metal backend. Whether performance improves can only be judged with separate measurement data.