ggml-org/llama.cpp is a repository that implements LLM inference in C/C++ and has collected 133629 stars, and the update bundled on November 7, 2026 contains 8 releases, from b12173 to b12189. The center of this batch is the strengthening of kernels across several backends and a change in the server's slot processing order.
First, the range of supported operations grew. The OpenCL backend supports repeat operations for the f16, i32, i16 and bf16 data types as of b12173, and added the arange operation in b12183. WebGPU gained a group norm kernel in b12187, and Metal fixed GROUP_NORM for the case where the batch dimension (ne03) is greater than 1 or the input is non-contiguous in b12189. b12186 added a readonly marker to the Vulkan decodeBuf declaration, and b12188 changed the order so that the server visits the parent slot before the waiting child slot. Beyond that, b12182 raised the computation-view limit for buffers with few tensors and reverted the DeepSeek 4.1 test model to 6 layers, while b12174 changed Windows resource files to be generated directly by CMake without a template input file.
That said, the changes carried in each build are generally limited to one or two kernels, or to the build and test path. For b12186 there is only a note that the performance gain is expected with future driver versions, with no measurement figures, and no performance or user metrics are presented either. In other words, this batch is close to a continuation of the work of widening, one by one, the kinds of operations usable on specific backends, and judging how much faster things became requires further per-backend measurements.