On Sep 30 2026 the open‑source llama.cpp library released version b11284. The update changes how OpenVINO handles weight views, fixing a GET_ROWS rejection that previously broke some quantized models. The fix folds row offsets into gather indices, so the view is treated as a regular constant. The release also lists a wide range of new build targets, from macOS Apple Silicon to CUDA 13 on Linux and Windows.
Why it matters
The change lets more quantized LLaMA models run on OpenVINO‑enabled devices without errors.
