Skip to content

llama.cpp Reaches v1.0.0 With a Stable Public C API

The local inference library and its ggml engine both hit v1.0.0, with semantic versioning now governing breaking changes.

Draftship Editorial3 min read

The release

On November 13, 2026, the llama.cpp team published llama.cpp v1.0.0 and ggml v1.0.0, the two C/C++ libraries behind the project's local inference stack. The announcement on the Hugging Face blog frames llama.cpp v1.0.0 as a statement about responsibility rather than a feature drop: after 3.5 years, many downstream applications depend on the library's interfaces, so the team says a flexible but stable API matters more than ever.

llama.cpp runs multi-modal LLMs on hardware ranging from consumer machines to enterprise servers, using quantization, optimized backends and a portable C/C++ implementation. The v1.0.0 label is about how that code is versioned and consumed.

What the version number now means

Both projects adopt semantic versioning, with rules the post spells out directly:

  • Major: a breaking change to the public C API under ./include.
  • Minor: a backward-compatible feature, backend update, new model support or API addition.
  • Patch: a bug fix with no API change.

The team says future major-version bumps will reflect API and ABI incompatibilities, and that it intends to be stricter about making those boundaries visible to downstream consumers as adoption grows. The public headers are llama.cpp/include/ for libllama and ggml/include/ for libggml. In practice, a major bump is the signal for anyone linking against those headers to review before upgrading.

Where it ships and what it runs on

The post lists the channels the projects reach developers through: pre-release artifacts for Linux, macOS, Windows, Android and iOS; Docker and Debian for Linux; Homebrew for macOS and Linux; conda-forge for Windows, macOS and Linux; MacPorts for macOS; Winget for Windows; and XCFramework for iOS and macOS. It points readers to an open discussion for adding ggml-org projects to packaging and release pipelines.

The engine underneath, ggml, is built around dedicated backends, each compiled as a dynamic library and optionally loaded at runtime. The team says llama.cpp currently supports more than 150 architecture variants and all mainstream speculative decoding methods.

On development, the post says CI capacity is being scaled with Hugging Face's own Buckets and Jobs products so that every change landing on the master branches is checked, backed by self-hosted runners on hardware from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel. The team also says it integrates LLMs into its own development cycle while keeping architectural control and implementation detail as priorities.

Limits and open questions

The announcement is a governance and distribution post, not a changelog. It does not describe any specific breaking change in this release, does not include a migration guide, and gives no benchmark numbers. It also does not say which backends are considered fully supported versus experimental, and it gives no dates for future releases. Readers who need to know exactly what moved between the last 0.x build and v1.0.0 will have to check the repositories.

What this means for building apps with AI

The practical effect for teams embedding local inference is predictability at the linkage layer: a pinned version stays pinned, and a major bump is now an explicit warning that headers or ABI changed. That is useful for anyone shipping a desktop or mobile product around a bundled model, though this suggests more than the post states about how quickly downstream projects will adopt the scheme.

Most people building web apps with AI will not link against these headers at all, since inference usually happens through a hosted API. If that is your path, Draftship's documentation on adding AI to your app covers how app-side AI features work on a hosted stack. Draftship is not affiliated with Hugging Face or the llama.cpp project, and nothing in the announcement concerns it.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.