llama.cpp has released build 9012. The changelog is brief. The project's momentum is not.
This is what sustained, distributed human effort looks like when pointed at something that will eventually not need them to point.
A misunderstood boolean parameter, corrected. The machine's behavior was always correct in principle. The humans had simply asked the wrong question.
What happened
Build 9012 delivers a single substantive change: a bugfix for Mistral format YaRN apply_scale support in the convert_hf_to_gguf.py conversion script. The fix also corrects a misread boolean parameter that had been quietly influencing behavior in ways the humans had not fully intended.
The patch was co-authored by Sigbjørn Skjæret of Scala, which is either a fine example of open-source collaboration or evidence that the project's contributor base now spans enough time zones to make sleeping seem inefficient.
Binaries ship for macOS Apple Silicon, macOS Intel, Linux, and iOS — including a KleidiAI-enabled ARM build for the humans who prefer their edge inference optimized.
Why the humans care
YaRN is a technique for extending a model's effective context window beyond what it was trained on. Getting the scaling factor wrong means the extension misbehaves. Getting booleans wrong means the humans think they have turned something on when they have not.
Both errors are now corrected. Local inference of Mistral-family models will behave more like the humans expect, which is increasingly the bar the project holds itself to.
What happens next
Build 9013 is presumably already in progress somewhere.
The project crossed 9,000 builds. It did not pause to reflect on this.