The open‑source llama.cpp library released version b11680 on Oct 16 2026. The update overhauls the Hexagon NPU code: it removes a hack, adds tensor repacking for F16/F32/BF16 weights, and supports tiled Q5_K get‑rows and raw k‑quant tensors. The changes also clean up many internal kernels and improve performance on several back‑ends, including CPU, CUDA, Vulkan and Hexagon‑enabled devices. The project lists builds for macOS, Linux, Android, Windows and several openEuler configurations.
Why it matters
Developers can now run llama.cpp models more efficiently on Hexagon‑powered phones and other specialized hardware.
