The ggml‑org team pushed the b12322 tag for llama.cpp on 12 Nov 2026. The change modifies how graph results reserve metadata, allocating only the needed pages instead of initializing the whole reserve. This reduces the amount of RAM the library touches during inference.
The update is available for all listed platforms, from macOS and Windows CPUs to Linux CUDA, Vulkan and OpenVINO builds.
Why it matters
Developers running LLaMA models on limited‑memory devices will see lower RAM consumption, enabling larger models or smoother multitasking.
