The ggml-org team released llama.cpp b12031 with initial OpenCL support for IQ3, two compact model-weight formats. The update adds operations for IQ3_S and IQ3_XXS, including matrix and vector operations, plus Adreno-specific paths for Qualcomm GPUs.
The release notes also describe decoding optimizations and support for narrower batches on Adreno. They provide no speed comparisons, so the practical performance change is not stated.
Why it matters
Developers running llama.cpp on Adreno GPUs can try these OpenCL paths with IQ3 models, but the release notes do not establish whether they run faster or which devices benefit most.
