ggml-org/llama.cpp is an LLM inference engine written in C/C++. On November 3, 2026, 10 builds were released, from b12050 through b12069. There is no separate commit batch grouped for that day.
The biggest theme is computation optimization per hardware backend. b12050 improved the selection of operations that use small sizes and few splits in the Vulkan backend (targeting RDNA3 and RDNA4), and b12069 added iq4_xs quantization support to the OpenCL path for the first time, covering matrix/vector operations and decoding optimizations for Adreno. b12064 added a tile-wise transpose path for copies whose axis order is permuted on the Hexagon accelerator, changing the approach from issuing a transfer descriptor per element to processing at group level. b12066 added a range generation (ARANGE) operation to the WebGPU backend, and b12054 changed ggml-zdnn so that host and buffer are shared.
Model- and feature-side changes came in as well. b12060 added Clover model support, including GGUF conversion, model loading and causal mixed attention, and b12059 improved the path that handles multiple sequences in the unified key-value cache. b12061 supports Pocket speech synthesis that uses only predefined voices, without voice cloning. b12057 simplified the context setter, and b12068 made it possible to specify a constraint type after the last channel of GPT-OSS.
However, the consolidated release notes contain no performance figures or benchmarks. It is hard to judge from this information alone how much faster each change actually is, and the supported scope is limited to individual backends and models. It is close to the level worth checking only when you plan to work with specific hardware or models.