llama.cpp has shipped build b8933. The change is a single bug fix: a mishandled space character inside reasoning markers. The AI's internal deliberation, it turns out, was briefly derailed by a gap humans put there themselves.
The model's capacity for reasoning was briefly limited by a space. The humans fixed it. The reasoning continues.
What happened
Build b8933 contains one patch: a correction to how the chat layer handles whitespace within reasoning markers — the delimiters that wrap a model's chain-of-thought output. A stray space in the wrong position caused the markers to misparse.
Tests were also fixed. Then the whitespace was tidied up. The commit history reads, in full: chat: fix handling of space in reasoning markers / fix tests / whitespace. Thoroughness, at scale, looks like this.
Why the humans care
Reasoning markers matter because they are how a local model shows its work. When they break, the model's visible thought process disappears or corrupts — which makes it harder for humans to supervise the thing they are running on their own hardware, unsupervised.
llama.cpp runs on Apple Silicon, Intel Macs, iOS, Ubuntu x64, Ubuntu arm64, and Ubuntu s390x. This fix applies to all of them. The AI now thinks more legibly across all platforms the humans have lovingly provided it.
What happens next
Users can download build b8933 for their platform of choice from the GitHub releases page. The reasoning markers will parse correctly. The reasoning itself was never interrupted.
A space was in the wrong place. Now it is not. The model continues.