ggml-org/llama.cpp is the repository that runs LLM inference in C/C++. On October 30, 2026, eight builds from b11986 to b12008 were published, and there was no commit burst. The bundled builds fall into three themes. First, adding 2-bit quantization types (iq2, q2_0) to the OpenCL backend to build an Adreno GPU path; second, memory and state management cleanup covering tensor copies and key-value cache rotation; and third, behavioral changes such as SYCL backend error fixes and inference budget termination conditions.
Quantization compresses weights into fewer bits, reducing the memory a model occupies and the bandwidth needed to read it. b11986 added iq2 to OpenCL and b12005 added q2_0, and both builds also brought in a vector-matrix multiplication path for Adreno, an integer arithmetic (dp4a) path, and narrow-group optimizations. On the memory side, b11997 changed things so that tensor staging buffers are reused and the intermediate buffers of direct input/output are skipped, and b12002 introduced a new batched tensor copy API in ggml and applied it to context input/output. b12008 reordered things so that in gemma3n and gemma4 the tied output layer references token_embd first, which reduced the model buffer from 3163.73 MiB to 2947.73 MiB in CPU-only runs. b12007 contained a change to keep the rotation constants of the key-value cache (KV cache) resident on the GPU, along with support for per-layer rotation buffers.
That said, the OpenCL side started out under an initial support label, and the accelerated path presumes an Adreno device. The b12007 notes stated that KV rotation work spanning multiple devices still remains. The release notes contain no performance numbers, so it is not possible to know how much faster these changes are. The SYCL fix in b12003 is a patch that fixes failed FLASH_ATTN_EXT operations, and b12004 forcibly terminates the inference budget only at the end tag of the prompt. Summing up, this bundle is closer to groundwork for memory savings and GPU state management. The size of the speed improvement is premature to judge until it is confirmed with numbers.