The ggml‑org team released version b11819 of llama.cpp on 2026-10-22. The update adds support for non‑standard GQA (grouped query attention) models and a BPE (byte‑pair‑encoding) tokenizer, plus a new domyn‑small pre‑type. It also adds metadata handling for Metaspace prepend schemes. The build matrix now lists many CPU and GPU targets, including Apple Silicon, CUDA 12/13, Vulkan, ROCm and Android Snapdragon.
Why it matters
Developers can now run a wider range of LLMs, including those using GQA, on more devices such as smartphones and GPUs.
