Skip to main content

llama.cpp 1.0 sets versioning rules and adds inference features

13 NOVEMBER 2026·2 MIN READ·2 SOURCES

ggml-org says it has released llama.cpp and ggml libraries v1.0.0, establishing versioning rules for future compatibility changes and adding features across inference, serving and distribution.

llama.cpp 1.0 sets versioning rules and adds inference features

Key takeaways · 4

  • 01

    Future API and ABI incompatibilities in llama.cpp and ggml will be signaled by major-version bumps, ggml-org says.

  • 02

    Breaking changes to the public C API trigger a major bump; compatible additions trigger a minor bump, and API-neutral bug fixes trigger a patch bump.

  • 03

    The release removes deprecated public API and promotes the NextN/speculative API into `llama.h`.

  • 04

    The blog lists pre-release artifacts for Linux, macOS, Windows, Android and iOS, as well as Docker for Linux and conda-forge for Windows, macOS and Linux.

A versioning contract for dependents

The release’s central commitment is about signaling compatibility changes: ggml-org says future API and ABI incompatibilities in llama.cpp and ggml will come with major-version bumps.[1] Under its semantic-versioning rules, a breaking change to the public C API triggers a major bump.[1] Backward-compatible features, backend updates, model support and API additions trigger a minor bump, while a bug fix that does not change the API triggers a patch bump.[1] For teams that build on these libraries, the rules offer a basis for planning around updates; they do not say that every update will be backward compatible.

API and inference changes

llama.cpp v1.0.0 removes the deprecated public API and promotes its NextN/speculative API into `llama.h`.[2] The release adds OCP FP8 key-value caches and a host-memory mixture-of-experts cache with a configurable CPU share.[2] It also adds support for loading external PLE shards, drama-3 text-to-speech, and vision support for MageVL-Cowork and Kimi-K3.1.[2] These additions sit within a project ggml-org describes as C/C++ LLM inference that provides the libllama library.[1] The evidence does not specify performance results for the new features.

Serving, security and distribution

The release hardens llama-server against SSRF, CORS issues, DNS rebinding and MCP tool-cancellation risks.[2] The server also gains an OpenAI-compatible `/v1/audio/speech` endpoint, and the release includes a redesigned web UI with chat groups and a markdown editor.[2] The blog lists pre-release artifacts for Linux, macOS, Windows, Android and iOS; Docker distribution for Linux; and conda-forge distribution for Windows, macOS and Linux.[1] ggml-org says it is expanding how developers and end users can obtain the software.[1]

Two libraries reach 1.0

The announcement covers both llama.cpp v1.0.0 and ggml v1.0.0; the release updates ggml to v1.0.0.[1][2] ggml is described as a tensor library for machine learning that provides libggml, while llama.cpp is described as LLM inference in C/C++ that provides libllama.[1] ggml-org says llama.cpp can run multimodal LLMs on hardware ranging from consumer-grade local machines to enterprise servers.[1] It attributes that reach to quantization, optimized backends and a portable C/C++ implementation, and says each backend can be compiled as a dynamic library and optionally loaded at runtime.[1]

For teams integrating local inference into products, explicit compatibility signals can help make upgrade planning more deliberate. The release also gives engineering teams changes to evaluate in APIs, caching, model support and server deployment, while the listed distributions may inform how they test installation paths.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 13 November 2026

    llama.cpp 1.0 sets versioning rules and adds inference features

Sources

AI fluency, one session a day, built for your work.