LLM inference in C/C++
- Language
- C++
- License
- MIT
- Stars
- 134.9K
- Try it
brew install llama.cpp
llama.cpp and its tensor library ggml were both tagged v1.0.0 on November 13, the first releases of the project to carry a formal semantic-versioning promise. The tag is live on GitHub: in the v1.0.0 source tree, CMakeLists.txt sets the llama.cpp version to 1.0.0 and ggml/CMakeLists.txt does the same for ggml.
Why it matters: llama.cpp is the engine underneath a long list of local-AI apps, wrappers and language bindings, and until now its version numbers told developers very little about whether an update would break their code. The project's announcement, as quoted by tbreak, says both libraries "will reflect API and ABI incompatibilities with major-version bumps." For anyone shipping software on top of libllama, that turns "read the commit log and hope" into "pin 1.x."
The rules, in one table
According to the release coverage from tbreak and explainx.ai, the policy applies across ggml-org projects:
| Version part | What it signals |
|---|---|
| MAJOR (1 → 2) | A breaking change to the public C API |
| MINOR (1.0 → 1.1) | Backward-compatible features, backend updates, new model support or API additions |
| PATCH (1.0.0 → 1.0.1) | Bug fixes with no API change |
The scope matters. explainx.ai notes the promise covers the C headers — include/ for libllama and ggml/include/ for libggml. It does not say that command-line flags, llama-server's HTTP interface, model outputs or every old GGUF file are frozen. If your setup is a shell script or an agent tool talking to a local server, the 1.x label alone doesn't protect you.
What was removed (this is the part that can break builds)
The 1.0 release clears out deprecated API before the promise starts. The release coverage says:
llama_set_warmupand 35 other deprecated functions are gone from libllama. We checked the v1.0.0llama.h:llama_set_warmupno longer appears.- ggml 1.0.0 prunes its own deprecated public API, according to explainx.ai and AI TL;DR.
- The
--defrag-tholdoption anddefrag_tholdsetting are removed, andclip.hhas been folded intomtmd, per explainx.ai. - The NextN / speculative-decoding API moved from
llama-ext.hinto the publicllama.h— the v1.0.0 header now has a "NextN / speculative decoding API" section — andllama_set_embd_state_srcwas renamedllama_set_ctx_state_src, according to explainx.ai; the new name appears in the v1.0.0 header.
Code that calls a removed function simply won't compile against 1.0.0.
What's new in the same release
The coverage lists a sizeable feature drop alongside the version change:
- FP8 (E4M3) KV cache support; the v1.0.0 headers include an
F8_E4M3type in both ggml and llama.cpp. - A host-memory cache for mixture-of-experts (MoE) expert weights, with a configurable CPU share.
- Loading per-layer-embedding (PLE) weights from separate GGUF shards.
- New model support including Kimi-K3.1-Pro, Ripple-65B-Preview, xing4_0 and the drama-3 text-to-speech model, plus vision for MageVL-Cowork and Kimi-K3.1.
- In
llama-server: an OpenAI-compatible/v1/audio/speechendpoint, cancellation of MCP tool calls when a client disconnects, and fixes for SSRF, CORS and DNS-rebinding issues.
AI TL;DR and tbreak both cite support for more than 150 model architecture variants. The repository had 134,925 stars when we checked GitHub's API on November 14, and it's MIT-licensed. Builds ship through GitHub pre-release artifacts, Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework for Apple platforms.
Who should do what
- You link libllama or libggml directly (bindings, embedded apps): build against v1.0.0 in a branch first, fix any removed or renamed calls, then pin a
1.xrange. - You only use
llama-cliorllama-server: check your flags (especially anything using--defrag-thold) and smoke-test your scripts. The C API promise doesn't cover them. - You use an app built on llama.cpp: nothing changes today. The benefit shows up over time if those apps can adopt updates with less breakage.
- Packagers: semantic versions mean you can hold back a future 2.0 while still taking 1.x fixes.
The real test comes later: whether the maintainers keep breaking changes out of minor releases on a project that has historically moved very fast. Watch for the first 1.1 and whether downstream projects start pinning version ranges.
Sources
- GitHub — ggml-org/llama.cpp at tag v1.0.0 (CMakeLists.txt, include/llama.h, ggml/include/ggml.h) github.com
- tbreak — llama.cpp 1.0 reaches stable release for local AI tbreak.com
- explainx.ai — llama.cpp v1.0.0: Semantic Versioning Comes to Local AI explainx.ai
- AI TL;DR — llama.cpp v1.0.0, first stable release, with semantic versioning ai-tldr.dev