The ggml‑org team released llama.cpp version b12363 on November 13, 2026. The update adds an optional device keepalive setting to the Vulkan backend, letting users specify how long the GPU stays active before shutting down. The change moves ownership of the device to the backend and simplifies the code. Pre‑built binaries for macOS, iOS, and several Ubuntu variants are included.
Why it matters
Developers can run repeated model calls on GPU without costly re‑initialization, improving response speed for local AI apps.
