Developers who run large models locally waited three and a half years: llama.cpp finally released version 1.0.0 on November 13. It is the first formal major release of this C/C++ inference engine, and the first time it has used semantic versioning since March 2023, when it started using raw commit numbers as version strings. The companion tensor library ggml shipped 1.0.0 at the same time, and both projects now promise that incompatible changes to the public C API will only appear in major releases.
What actually changed in 1.0.0
The headline change is a cleanup. Version 1.0.0 removes 35 deprecated public functions, along with old options such as the warmup setter and the defragmentation threshold, while promoting the NextN speculative-decoding API from an extension header into the main llama.h header, bringing it under the protection of the new versioning promise. New capabilities include OCP FP8-format KV caches, a host-memory expert cache for mixture-of-experts models, and loading external PLE or engram data from separate GGUF shards. According to the project, llama.cpp now supports more than 150 model architecture variants and all mainstream speculative decoding methods.
Who this upgrade affects
The direct impact falls on tools built on llama.cpp, most notably local AI apps such as Ollama and LM Studio: they can finally rely on a stable public API instead of holding their breath every time they follow a new commit number. For ordinary users who only run ready-made apps, the version number itself changes nothing visible; the benefit is that future upgrades become more predictable and compatibility problems rarer. One warning, though: 1.0.0 is itself a cleanup release, so any app using a removed API must update its code before upgrading — it cannot just swap in the new binary. The bar for local deployment is shifting, as shown by the recent example of Edge0 squeezing a 145B model onto a laptop.
Stable does not mean never changing
The semantic versioning promise is specific: major versions cover incompatible changes to the public C API, minor versions add features without breaking compatibility, and patch versions fix bugs only. But it governs the public C API, not model outputs, performance, or every command-line option. The project also lists complete distribution channels: prebuilt artifacts cover Linux, macOS, Windows, Android and iOS, with Docker, Debian, Homebrew, conda-forge, MacPorts and Winget alongside. The local AI ecosystem has moved from 'it runs' to 'the interface comes with a promise' — that is what 1.0.0 is really about.
