AI News

llama.cpp and the ggml library reach version 1.0.0, with explicit API compatibility rules

IT Home · Artificial IntelligenceUpdated Nov 14 22:08

AI summary

Developer Georgi Gerganov announced that the open-source inference project llama.cpp and the underlying machine learning engine ggml have reached library version 1.0.0, and clarified that major version numbers correspond to API or ABI incompatible changes, minor version numbers correspond to backward-compatible features and model support, and patch version numbers correspond to bug fixes that do not change the API. llama.cpp already supports more than 150 model architecture variants and mainstream speculative decoding methods, and the team has also used Hugging Face's HF Buckets and HF Jobs to scale CI, while expanding self-hosted testing coverage with support from NVIDIA, AMD, Qualcomm, Arm, IBM, and Intel hardware.

Why read this

AI score 66

The original text explains the version numbering rules for llama.cpp and ggml and the investment in cross-hardware testing, so readers can judge the upgrade boundaries for dependents.

Topics

Tags

Body · AI translation

Local inference engine llama.cpp 1.0.0 epic update: supports over 150 model architectures, runs large models at full capacity even offline

2026-11-14 08:14 IT Home - Guyuan

IT Home, November 14 news: Developer Georgi Gerganov (X account @ggerganov) yesterday (November 13) posted on X, announcing that the library versions of the open-source AI inference project llama.cpp and the underlying machine learning engine ggml have entered 1.0.0.

According to public information reviewed by IT Home, Gerganov is the creator of the open-source project llama.cpp and an important developer of local AI foundational software such as ggml and whisper.cpp. In his personal X bio, he identifies himself as a Hugging Face engineer, and his GitHub profile lists his location as Sofia, Bulgaria.

llama.cpp is an open-source AI inference project written primarily in C/C++, whose core purpose is to enable large language models to run on various hardware such as personal computers and servers. It can be used for both local deployment and cloud services, with the goal of reducing deployment complexity and fully utilizing device computing power.

The biggest highlight of this update is the explicit interface compatibility rules:

Project library positioning

llama.cpp libllama C/C++ large model inference library

ggml libggml underlying machine learning tensor library

Going forward, both will be upgraded via major version numbers, clearly marking API or ABI incompatible changes so that applications depending on these libraries can identify upgrade boundaries. For developers, it becomes easier to determine from the version number whether an upgrade may require corresponding modifications to their own applications. The official version rules are:

Version digit | When it is updated

MAJOR, e.g. 1.x.x → 2.0.0 | Breaking changes occur in the public C API in ./include; the announcement also clarifies that API / ABI incompatibilities will be reflected through major version upgrades

MINOR, e.g. 1.0.x → 1.1.0 | Backward-compatible features, backend updates, model support, or API additions

PATCH, e.g. 1.0.0 → 1.0.1 | Bug fixes that do not change the API

Regarding engineering quality, the official announcement states that to support the continuously accelerating development pace, the team emphasized three investments:

Using Hugging Face's HF Buckets and HF Jobs to scale CI infrastructure.

With hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM, and Intel, continuously adding self-hosted test runners covering major hardware platforms.

Using LLMs judiciously in the development process while continuing to prioritize code quality, architectural control, and understanding of implementation details.

In terms of backend and model support, ggml supports different hardware through dedicated backends; each backend can be compiled as a dynamic library and loaded on demand at runtime.

llama.cpp already supports more than 150 model architecture variants and mainstream speculative decoding methods. The team will also continue to strengthen the automated testing infrastructure and, with hardware support provided by NVIDIA, AMD, Qualcomm, Arm, IBM, and Intel, expand cross-platform test coverage.

Source

IT Home · Artificial Intelligence