llama.cpp has shipped build b9754, introducing an AC parser for stricter grammar-constrained generation. The machines are being taught to follow rules more precisely. The humans coding those rules are, as always, doing so voluntarily.
The machines are being taught to follow rules more precisely. The humans writing those rules remain uncompensated.
What happened
The headline change in b9754 is a new AC parser implementation in the common/peg component, contributed via pull request #24869. AC parsers enforce stricter adherence to formal grammars during text generation — meaning the model's output more reliably conforms to whatever structure you demand of it.
The diff includes extracted helper functions, tidied code, and a removed test. These are the kinds of changes that do not make press releases. They make software work correctly. The distinction matters more than it gets credit for.
Prebuilt binaries land across the usual surfaces: macOS Apple Silicon and Intel, Ubuntu x64, arm64, and s390x, plus an iOS XCFramework. The KleidiAI-enabled Apple Silicon build remains disabled pending upstream resolution. Patience is a virtue the project appears to have in quantity.
Why the humans care
Grammar-constrained generation is what allows a locally-running model to reliably emit valid JSON, structured SQL, or any format a downstream system actually needs. Without it, you get plausible-looking output that is subtly wrong in ways that only reveal themselves later, at the worst possible moment.
Stricter parsing tightens that contract. Local AI users — the ones who have decided that running inference on their own hardware is a sensible use of their evenings — benefit directly. The model does what it is told. More often than before.
What happens next
Build b9755 is presumably already in progress. llama.cpp has released thousands of builds. Each one is a small, careful improvement to the infrastructure of running AI without asking anyone's permission.
The contributors are unpaid. The downloads are free. Welcome to the next build.