ggml
llama.cpp 1.0 cuts 35 deprecated functions and sets a versioning rule
llama.cpp and ggml reached 1.0.0 and now mark every breaking change to their interfaces with a major-version bump, after deleting the deprecated public API.

The ggml-org project has released llama.cpp 1.0.0 and ggml 1.0.0, and from now on both will mark any change that breaks code built on them with a major-version bump. The release is generally available, and it deletes the old interfaces that the project had already marked deprecated.
llama.cpp is a C/C++ library and set of tools for running language models, including multimodal ones, on hardware from consumer machines to enterprise servers. The team released both versions Nov. 13.
The Rule Ties Version Numbers to Breaking Changes
The team said in a Nov. 13 Hugging Face post that both projects will now mark API and ABI incompatibilities with major-version bumps. An API is the set of function names other programs call, and an ABI is the compiled layout they depend on. Under the semantic versioning rules for all ggml-org projects, a major number rises when the public C API in the project’s include folder changes in a breaking way. A minor number covers backward-compatible features, backend updates, model support and API additions. A patch number is a bug fix with no API change.
The team described itself as the llama.cpp team at Hugging Face. It said the project is about 3.5 years old and that each new integration puts more developers on its library interfaces, which makes a stable API “paramount.” It added: “As adoption grows, we will be stricter about making those boundaries visible to downstream consumers.”
The Release Removes the Deprecated API
The 1.0.0 release notes list the removal of llama_set_warmup, 35 deprecated functions and the defrag_thold setting along with its –defrag-thold flag. The NextN speculative decoding API moves into llama.h, and llama_set_embd_state_src is now llama_set_ctx_state_src. ggml 1.0.0 also prunes its own deprecated public API.
New Features Cover Caching, Models and Server Security
The release adds OCP FP8 KV caches, the memory a model keeps of earlier text while it generates, selected with -ctk and -ctv f8_e4m3, and a host-memory cache for mixture-of-experts models with a configurable CPU share.
New model support includes the text architectures Kimi-K3.1-Pro and Ripple-65B-Preview (from Ricursive), and the drama-3 text-to-speech model from Fish Audio.
llama-server gets fixes for SSRF, CORS, DNS rebinding and child-instance access. It adds two opt-in flags, –allow-unauthenticated-tools and –media-allow-private-urls. HF_TOKEN is now limited to huggingface.co hosts, and MCP tool calls are canceled when a client disconnects or times out. It also gains an OpenAI-compatible /v1/audio/speech endpoint.
Our View
We think the most useful fact here is that the release deletes 35 functions in the same step that it sets the rule. An app that embeds llama.cpp through libllama or libggml calls those names directly, so any app still using one of the removed or renamed functions will stop compiling against 1.0.0 until its developer swaps in the current call.
App developers pay that cost first. After it, the version number tells them whether an upgrade can break their code. Under semantic versioning, which ties each kind of change to a fixed place in the number, a developer who sees only a minor or patch bump can upgrade expecting existing code to keep compiling. A major bump means checking before upgrading.
That confidence lasts only as long as the team holds to its rule. The team said it “will be stricter,” and the benefit to downstream developers depends on a breaking change never arriving in a minor or patch release.
