ggml-org/llama.cpp is a project that performs large language model (LLM) inference in C/C++. Build b11976 has been published in the repository. According to the release summary, this change concerns the way im2col is handled in the Hexagon backend. im2col is the name of the operation used in the change description, and the release notes state that this work is now split by output column blocks and assigned to multiple threads. This content is based on the b11976 release record published on October 29, 2026.
The scope of the change is the way im2col is processed in the Hexagon backend. Splitting by output column blocks means dividing the processing work according to groups of output columns, and the use of multiple threads is the concrete difference seen in this build. So the news in b11976 is not an addition of supported models or user-facing features, but the fact that a specific operation within that backend has been changed to be processed in parallel. The repository description introduces llama.cpp as a C/C++-based LLM inference project, and the release summary technically limits this change to the Hexagon backend.
However, the provided release summary contains no comparative figures for processing time or throughput. It is therefore not possible to judge from this information alone how much the change to split the work across multiple threads actually affects inference speed. What can be confirmed is only that, in b11976, im2col processing in the Hexagon backend changed to a multi-threaded approach based on output column blocks. The magnitude of the performance improvement cannot be concluded from the facts presented.