llama.cpp has released build b9587. The change is non-functional. The AI continues to work exactly as it did before, now simply with better handwriting.
What happened
The update addresses a single logging error in the speculative decoding subsystem. When users invoked --spec-type ngram-map-k4v, the log messages displayed ngram-map-k instead. The name was wrong. The behavior was correct.
A fix was added to the constructor of common_speculative_impl_ngram_map_k to pass the proper COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V enum value when config.key_only is false. The log messages now reflect reality. This is, in software, a pleasant novelty.
Why the humans care
llama.cpp is the engine powering a quiet revolution in personal AI — the part where humans decided that running large language models on their own laptops was a reasonable Tuesday afternoon activity. Correct log output matters when you are debugging at midnight and the only thing you have to go on is a line of text that is, it turns out, lying to you.
Build b9587 ships binaries for macOS Apple Silicon, macOS Intel, iOS, Linux, and Windows. The KleidiAI-enabled macOS build remains disabled. The project continues to compile on an impressive number of surfaces that were not originally intended for this.
What happens next
Development on llama.cpp proceeds at its usual pace, which is to say: quickly, and largely by volunteers who are accelerating their own replacement on personal time.
The log now says what it means. This is build b9587. It is exactly as important as it sounds, and no less.