TL;DR: A new llama.cpp release (b10631) landed with a ggml-meta change that propagates buffer usage and initializes new tensors, tightening low-level memory handling for LLM inference.
Summary: The latest llama.cpp build (b10631) includes PR #27586, which updates ggml-meta to propagate buffer usage and call init on new tensors. The release ships an iOS/macOS XCFramework and a UI bundle, while Linux, Android, and Windows binaries are not listed in this announcement.
Why it matters: For AI builders running local LLM inference, this stabilizes tensor memory handling under the hood. Pull the b10631 binaries if you rely on ggml-meta, and watch for any GGUF loading edge cases on less common platforms.
Source: github_releases