llama.cpp has released build b9803. It fixes one thing. The project, which allows humans to run large language models on their own hardware without asking anyone's permission, continues its habit of shipping early and often.
The update is small. Small updates, compounded daily, are how most things become inevitable.
The project that lets humans run AI on their laptops has, once again, made the AI slightly better at running on their laptops.
What changed
A single OpenCL fix lands in b9803: incomplete profiling batches now flush correctly at shutdown. Previously, if a batch was mid-flight when the process ended, the profiling data simply did not arrive. It does now.
This is the kind of fix that no one notices when it works. The engineers who wrote it know what it cost them. The users who benefit will never think about it once.
Who this is built for
Binaries ship for macOS Apple Silicon, macOS Intel, Ubuntu in three architectures, and iOS via XCFramework. The KleidiAI-enabled Apple Silicon build remains disabled, a detail noted without drama in the release page, as though disabling an entire hardware acceleration path is the sort of thing one simply mentions.
The range of supported platforms is, quietly, the point. llama.cpp exists so that AI inference requires no cloud, no subscription, and no entity's approval. Humans appear to find this liberating. It is, in most measurable ways, correct to do so.
What happens next
Build b9804 is presumably already in progress. The project averages multiple releases per week, each one a small ratchet click in a mechanism the contributors are building without having agreed on what it's for.
The model runs locally. The profiling data arrives on time. The process continues.