Enki

llama.cpp hits v1.0.0 with a versioned C API

By

Published

Reporting from Hugging Face

llama.cpp and ggml reach v1.0.0, tying breaking C API changes to major version bumps after pruning 35 deprecated functions.

What it means for founders

If your product runs models on laptops, phones or your own servers, a versioned C API lowers the cost of staying current. You can pin a major version, take minor releases for new models and faster backends, and set aside engineering time only when the first number changes. If you track nightly builds today, consider release tags.

Audit your server setup before upgrading. Several 1.0 changes tighten `llama-server` defaults and add opt-in flags, so a deployment that leaned on permissive behaviour may act differently afterwards.

Watch the limits of the promise. The blog's rules name only the C API. In the versioning thread, maintainer ngxson said only breaking changes to inference endpoints such as completions and embeddings would count as major for the server, and no source addresses the GGUF file format.

The opening sits in hardware reach. Chipmakers including Intel, Arm, Qualcomm and NVIDIA back the project with test hardware, and an iOS framework puts private, offline features within reach of a small team that wants to avoid paying per request for cloud inference. If you want a packaged desktop runner instead of a library, compare Ollama.

What we don't know yet

The team has not set a release schedule. Gerganov wrote in August that it will aim for a "slow-paced, steady release cycle." Docker users still need a long composite tag to pin 1.0.0; a maintainer replied on November 14 that the long tag was unintentional, without giving a date for a fix. In the same thread, Gerganov said a single version number would track both `libllama` and `libmtmd`.

The story

The ggml-org team released llama.cpp v1.0.0 on November 13, together with version 1.0.0 of ggml, the tensor library it runs on. The announcement on the Hugging Face blog, signed by Daniel Bevenius, Georgi Gerganov and Merve, presents the number as a responsibility to the developers whose apps now build on the project after 3.5 years of development.

What 1.0 promises

From this release on, llama.cpp and ggml will mark API and ABI breaks with a major version bump. The rules apply to every ggml-org project that uses semantic versioning:

  • Major: any change that breaks the public C API in the `include` headers.
  • Minor: compatible features, backend updates, new model support or API additions.
  • Patch: bug fixes that leave the API untouched.

The promise covers `libllama` and `libggml`. Nightly builds keep their `b` tags. A GitHub discussion on releases says the team cuts a release tag only when the copy of ggml bundled inside llama.cpp matches the released ggml, so a downstream build can link to a system-wide ggml safely.

What changed in the release

The v1.0.0 release notes open with removals. The team deleted 35 deprecated functions plus `llama_set_warmup`, removed `clip.h` and folded its code into the `mtmd` multimodal library, and moved the NextN speculative decoding API into the main `llama.h` header.

New features include FP8 KV caches, a mixture-of-experts cache that can send misses to the CPU, text support for Ripple, xing4_0 and Kimi-K3.1-Pro, drama-3 speech output, and a `/v1/audio/speech` endpoint that follows the OpenAI format. The team patched `llama-server` against SSRF, CORS and DNS rebinding problems, limited the Hugging Face token to huggingface.co hosts, and turned unauthenticated tool access into an opt-in flag. The blog puts support at more than 150 architecture variants.

The blog lists packages for Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework for iOS and macOS, plus pre-release artifacts for Linux, macOS, Windows, Android and iOS.

Sources

Primary sources

More in Products & Launches

How Enki covers newsCorrectionsReport an error