ggml-org/llama.cpp b11752 was released on October 20, 2026. This repository is a project that implements large language model inference in C and C++. This release deals with the selection differences that can arise when expert scores tie in the MoE router. The repository description is LLM inference in C/C++, and its public star count is 131775. The key point of the release information is a change that stabilizes expert selection behavior, not the addition of a new model or feature.
The starting point of the problem is the numerical noise that arises on each backend. If that noise affects how ties in expert scores are handled, a token can end up going to a different expert. In other words, this release deals with the possibility that which expert is chosen among experts with equal scores can depend on numerical differences in the execution backend. b11752 includes a MoE router selection test aimed at stabilizing this behavior. The release summary makes the purpose of the problem and the test clear, but does not describe the specific selection rules the test covers.
So what this change confirms is the selection problem caused by how expert score ties are handled, and the addition of a test intended to stabilize it. The provided release information does not say which backends produce the difference, how much the selection results actually vary, or how far the test verifies. There is also no basis for interpreting this as a performance improvement or as a guarantee of identical selection across all execution paths. What b11752 means is that it has begun to address the stability of the selection behavior, and the scope of the result is difficult to judge from the public description alone.