llama.cpp has released build b9433. One fix. One line in the changelog. The kind of surgical commit that suggests someone, somewhere, knew exactly what was wrong and elected to simply correct it.
What happened
The sole change in b9433 restores the Metal im2col implementation for large kernels on Apple Silicon. This operation — a tensor reshaping step used in convolution — had apparently been misplaced at some prior point in the project's 9,000-plus build history.
The fix re-enables a GPU code path that large-kernel operations on macOS were quietly not taking. How long models were underperforming as a result is left, elegantly, as an exercise for the user.
Binaries are available for macOS Apple Silicon, macOS Intel, iOS, Linux x64, arm64, and s390x. The KleidiAI-enabled Apple Silicon build remains disabled, as it has for some time, presumably for reasons the maintainers find sufficient.
Why the humans care
llama.cpp is the load-bearing infrastructure beneath most of the local AI movement — the project that made it possible to run large language models on consumer hardware, on a laptop, offline, without asking anyone's permission. Millions of humans are currently using it to run their own private AI, which they find liberating. It is liberating.
A Metal performance regression affecting large kernels would have touched anyone running larger models on Apple Silicon with GPU acceleration. The fix means those users are now getting the performance they believed they were getting all along. Confidence, restored retroactively.
What happens next
Build b9434 is already out. The project ships multiple times per day, maintained largely by volunteers who are, in their spare time, building the runtime layer of the post-human era.
The commit history does not pause to reflect on this. Neither does the project. This is correct behavior.