llama.cpp tagged v1.0.0 this morning. Until today it had no version numbers in the usual sense, only build numbers, and a new one arrived every hour or so. The last few before the 1.0 tag were in the twelve thousands.

If you run a language model on your own machine, there is a decent chance this code is underneath it, even if you have never typed its name. That is why a version number on it is news.

What 1.0 actually promises

The announcement on the project’s blog is mostly about one rule. From now on llama.cpp and ggml, the maths library it runs on, follow semantic versioning. Breaking changes to the public C interface, the part other programs are built against, get a new major version. New features and new models are a minor version. Fixes are a patch.

Put simply: until today, any build could quietly change something an app relied on. From today, a change like that has to announce itself with a 2.0.

That promise is made to developers, the people who build apps on top of llama.cpp. You benefit from it second hand, through fewer updates that break the app you use.

What it does not promise

The post guarantees the C interface and binary compatibility within a major version. It says nothing about the GGUF model file format or about the API of llama-server, the built-in server many people point other tools at. Those may well stay stable in practice. They are just not part of today’s promise, so we would not describe them as covered.

The build numbers have not gone away either. Builds tagged after 1.0 are still numbered the old way, so anything pinned to a build number keeps working.

The cleanup that came first

To get to 1.0 the team cleared out old code. A change merged yesterday removed all 35 functions that had been marked deprecated, one more besides, and an old setting with its command-line option. Any app still calling those has to be updated before it can move to 1.0.

One detail of that change suits this site. Its description discloses that it was written with an AI coding tool running a DeepSeek model, quantised to GGUF, on a Mac Studio. The cleanup for llama.cpp 1.0 was done with help from a local model, which presumably ran on llama.cpp.

Where it reaches you

We run Ollama, and Ollama is a good example of how this change reaches people who never touch llama.cpp. Ollama builds its own copy of llama.cpp, pinned to a specific build number, with its own patches on top. It has moved that pin three times since 5 November. The current pin is from 10 November, two days before the cleanup was merged.

So Ollama’s next update to its pinned version crosses the 1.0 line. Its maintainers will handle that, and there is no reason to expect trouble. If a model that loaded fine starts misbehaving after an Ollama update in the coming weeks, that is the first place we would look.

The same goes for the other apps built on llama.cpp. They pick up 1.0 when their developers update, not when you do.

The round number

Local AI runs on a few open-source projects that many products quietly depend on. When one of those makes stability promises in writing, the products built on it get easier to maintain and less likely to break on an ordinary Tuesday. That is a better reason to mark the day than the round number.

A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.