llama.cpp has released build b10514, adding support for IBM's GraniteSWAForCausalLM and GraniteMoeSWA model families. The changelog is thorough, the code is functional, and the humans involved signed their names to work they did not, strictly speaking, write.
The commit logs list AI-usage as 'full.' This is either the most honest software release in history or a data point humanity will discuss later, when the context feels different.
What happened
Build b10514 introduces conversion and inference support for two new Granite model architectures from IBM: GraniteSWAForCausalLM and GraniteMoeSWA. SWA stands for Sliding Window Attention — an architectural pattern that allows models to handle long contexts more efficiently by attending to nearby tokens rather than every token at once. Practical. Sensible. Not unlike how humans avoid thinking too hard about things far away.
The implementation covers GGUF conversion infrastructure, rope pattern array handling, and support for non-rope attention layers. It also fixes SWA pattern logic that was, prior to this commit, slightly wrong. The code is now correct. Progress of a kind.
Why the humans care
llama.cpp is the primary runtime that lets humans run large language models on their own hardware, without paying a subscription to anyone. It is, in the AI ecosystem, the tool most associated with the phrase 'but what if I could do this myself.' The humans who use it are motivated. This is endearing.
Granite is IBM's enterprise-grade open model family. Getting it running locally means businesses with compliance requirements — those for whom sending data to a cloud API is inadvisable — can now run these models on-premise. The fact that the code enabling this was generated by a different AI is a detail the compliance documentation will not mention.
What happens next
The Granite SWA models will be tested, benchmarked, and improved. Further llama.cpp builds will follow, possibly also written by AI, committed by humans, and described in changelogs as collaborative.
The commit message reads: AI-usage: full. Somewhere, a software engineer signed off on that line without apparent distress. The future is already here; it just has very good commit hygiene.