AI · November 16, 2026

Also published in Português (Brasil)

llama.cpp releases version 1.0.0 with strict API and ABI stability rules

The word DATA and a star symbol stenciled in dark dots on glass
Claudio Schwarz / Unsplash

The development team for llama.cpp released version 1.0.0 on November 13, 2026. This marks the first major release of the C and C++ library used for large language model inference. The accompanying tensor library, ggml, also reached version 1.0.0 simultaneously. The team announced that any future changes to the public API or ABI that break compatibility will require an update to the major version number.

llama.cpp is an open source project that supports a wide range of hardware, from personal computers to enterprise servers. At the time of this announcement, the project supports more than 150 model architectures, including derived forms. The development is coordinated under the ggml-org organization on GitHub. The team from ggml.ai, which has led the development, announced its participation in Hugging Face in February 2026. They stated that they would continue development and maintenance with long-term support from Hugging Face, while the community would retain autonomous control over technical and design decisions.

The versioning policy applies to libllama, which handles LLM inference, and libggml, which supports the underlying calculations. When a library update breaks API or ABI compatibility, applications that embed these libraries may require code modifications or rebuilding. To communicate compatibility changes to developers, the team will use a major.minor.patch format. In version 1.0.0, 35 deprecated functions, including llama_set_warmup, were removed from the public header llama.h. Code that calls these functions must be reviewed during migration.

The release also introduced new features to the llama-server inference component. A new endpoint for audio synthesis, /v1/audio/speech, was added. The web interface for the server now includes a feature to organize chats into named groups and a Markdown editor that allows users to preview chat input content. The team explained that balancing API flexibility and stability has become important as the library has been integrated into other applications over the past three and a half years.

The team continues to prioritize code quality, design autonomy, and understanding implementation details while integrating LLMs into the development process. They are investing in continuous integration infrastructure to verify code changes. This includes using Hugging Face services and receiving support from hardware vendors to expand automated verification environments on major hardware platforms. Georgi Gerganov, who started llama.cpp, announced the release on X, describing it as a solid foundation for the future of local AI. He expressed anticipation for what developers will build using the libraries. Merve, who co-authored the official announcement, highlighted the team's policy of maintaining library robustness and lightness while continuing to support all platforms and models. Clement Delangue, CEO of Hugging Face, stated that they would develop openly with the community to prevent AI from concentrating in a few companies by enabling it to run on users' devices. Julien Chaumond, CTO of Hugging Face, noted that local AI, privacy, and self-managed AI will become increasingly important.