Money and business in the Middle East.

Economy

llama.cpp hits 1.0.0, promising stable API for local AI apps

ggml-org released llama.cpp and ggml 1.0.0 on November 13, 2026, adopting semantic versioning.

· Last verified: 14 Nov 2026 (ggml-org on Hugging Face, Unite.AI, DEV Community, tbreak)

Summary

  • ggml-org released llama.cpp and ggml 1.0.0 on November 13, 2026, adopting semantic versioning.
  • The projects ran on rolling 0.x numbering for roughly 3.5 years before the switch.
  • Downstream apps running models on laptops and phones now get advance warning of breaking changes.

The latest

The C/C++ inference library behind much of the local-AI ecosystem now carries a formal stability contract. ggml-org said llama.cpp and its underlying tensor library ggml both reached 1.0.0 on November 13, 2026, moving to semantic versioning after roughly 3.5 years of rolling 0.x builds. Any break in the public C interface will now be signaled by a major-number bump.

Details

  • The versioning rule: Under the new policy, the major number changes only for breaking changes to the public C interface in the include directory; the minor number covers backward-compatible additions such as a new feature, backend update, model family or interface; and the patch number covers fixes that leave the API untouched.
  • API and ABI: The blog said both projects will reflect API and ABI incompatibility through an increase in the major number. The ggml repository moved from build-style numbers such as 0.9.0-dev to semantic versioning in a commit titled ggml : introduce semantic versioning.
  • Release discipline: According to Unite.AI, no official llama.cpp tag is created unless the internal ggml version inside it matches the published ggml release, a rule aimed at keeping Debian and Homebrew packages consistent with upstream.
  • The cleanup: Unite.AI's reading of the release notes says 35 deprecated functions were removed from the public interface, while the NextN speculative decoding interface moved out of llama-ext.h into the public header llama.h.
  • Memory changes: The release adds an FP8 key-value attention cache and a host-memory buffer for mixture-of-experts models with a configurable CPU share, according to Unite.AI's reading of the release notes.
  • New model support: Text support was added for Kimi-K3.1-Pro, Ripple and xing4_0, alongside drama-3 for text-to-speech. Vision support covers MageVL-Cowork and Kimi-K3.1. The project lists more than 150 supported architecture variants.
  • Server security: The update includes fixes for server-side request forgery, CORS and DNS rebinding issues, restricts HF_TOKEN to huggingface.co hosts, and adds an OpenAI-compatible /v1/audio/speech endpoint.
  • Interface work: The web UI gains named conversation groups and a Markdown editor with live preview. Mobile builds add device safe-area layouts, full-screen windows, and an Android chat editor that opens the virtual keyboard.
  • Distribution reach: Prebuilt binaries cover Linux, macOS, Windows, Android and iOS, distributed through Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework for iOS and macOS. Self-hosted test runners use hardware from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.
  • What stays: Build numbers continue alongside the new scheme, with b12207 dated November 8 still listed on the releases page. The blog points to the Llama desktop apps for macOS and Windows without stating how 1.0 affects them, saying "a flexible but stable API is paramount."

Background

llama.cpp is the C/C++ engine that runs large language models directly on consumer hardware, from laptops to phones, without a cloud service. ggml is the tensor library underneath it, which ggml-org describes as designed to run on virtually any existing hardware.

Between the lines

The packaging rule tying llama.cpp tags to matching published ggml versions points at the real audience for this release: distribution maintainers in Debian and Homebrew, not end users. Removing 35 deprecated functions in the same release that promises stability suggests the cleanup was front-loaded deliberately, clearing legacy surface before the major-number commitment makes removals expensive.

What's next

Watch whether the next llama.cpp tag holds to the matched-ggml-version rule, and whether the FP8 KV cache and MoE host-memory buffer reach downstream packages in Debian, Homebrew and conda-forge.

Source: ggml-org on Hugging Face, Unite.AI, DEV Community, tbreak

llama.cpp hits 1.0.0, promising stable API for local AI apps · INXEN