Skip to content

llama.cpp reaches 1.0 and finally commits to semantic versioning

After three and a half years the local inference engine locks its public API, strips 35 deprecated functions and hardens its server against SSRF and DNS rebinding.

Source: ggml-org

llama.cpp, the C/C++ engine that most local-LLM tools quietly run on top of, shipped v1.0.0 on 13 November. It is the project’s first release under explicit semantic versioning: a major bump now means a breaking change to the public C API in ./include, minor covers new features and model support, patch is bug fixes only.

Getting there meant removing 35 deprecated functions, including llama_set_warmup and the --defrag-thold flag, and promoting the NextN speculative-decoding API out of the llama-ext.h extension header into the main llama.h. The server also gets real hardening: protection against DNS rebinding and CORS-proxy SSRF, new --allow-unauthenticated-tools and --media-allow-private-urls opt-ins, and the HF_TOKEN restricted to huggingface.co hosts only. It adds an OpenAI-compatible /v1/audio/speech endpoint and a redesigned web UI, and now covers more than 150 model architectures.

Why it matters: if you build on llama.cpp directly, or through something like Ollama that wraps it, you have been exposed to breaking changes on any release for over three years. That stops now: minor and patch releases are safe to pull automatically, and you only need to read the changelog closely when the major version moves.

The caveat: removing those 35 functions is itself a breaking change. If you are upgrading from an earlier build and called any of them directly, check the release notes before you bump.

Share

Related reading

More from The Wire

All briefs