The llama.cpp team at Hugging Face has released llama.cpp 1.0.0, the first stable version of the libraries for running models locally. Now the version number tells developers whether an update can break their application. Incompatible API and ABI changes will bump the first number, for example from 1.0.0 to 2.0.0. The second number marks backward-compatible additions, and the third marks bug fixes with no API change. This scheme is called semantic versioning (semver). It lets teams prepare and test the necessary changes in advance and lowers the risk that an AI chat or another service stops working for users after an upgrade, which makes updates easier to plan. Version numbers existed before 1.0.0. This release commits the project to compatibility guarantees.
The new version also adds more compact storage for intermediate computations, new text-to-speech features and extra server protection. ggml, the component that does the math, was updated alongside llama.cpp. The two projects ship two libraries: libllama and libggml.
Moving to 1.0.0 may itself require changes. The team removed 35 deprecated functions and, separately, llama_set_warmup, the function that configures model warmup (preliminary computations). Part of the image and audio code was also restructured: it was consolidated into a single component, mtmd. Applications that used the affected features will need to be adapted.
What changed in memory handling and models
The KV cache now supports FP8. The team also improved how MoE models run: their main data can stay in system RAM, while frequently used experts are also kept in GPU memory. If a required expert is not there once the cache is full, part of the computation can be handed to the CPU instead of transferring the data to the GPU. This can speed things up, but the result depends on hardware and settings.
Newly supported models include Kimi-K3.1-Pro, Ripple-65B-Preview and xing4_0 for text, drama-3 for speech generation, and MageVL-Cowork and Kimi-K3.1 for image processing. In total, according to the team, llama.cpp supports more than 150 architecture variants.
Voice features and server security
The project includes llama-server, which accepts requests from a browser or an app and passes them to the model. It now has an OpenAI-compatible text-to-speech endpoint, /v1/audio/speech. This makes it easier to plug local speech generation into applications that already use this request format.
The browser interface gains chat groups, a live preview of message formatting and a better layout on phones.
Protection for home and office networks has been strengthened as well. Running tools on a server that other devices can reach without an API key now requires explicit opt-in (--allow-unauthenticated-tools). Loading images and other media from the internal network needs explicit permission too (--media-allow-private-urls).
The Hugging Face access token is no longer sent automatically to arbitrary model download addresses. By default it goes only to huggingface.co, and another trusted address can be set manually.
Where to get the update
The developers list eight distribution channels, from prebuilt binaries on GitHub to package managers and frameworks for Apple apps. You need to check each channel separately to see whether it already has version 1.0.0.
As a Debian packager noted in the project’s discussion, predictable compatibility makes it easier to distribute llama.cpp. Official releases also align the versions of llama.cpp and ggml, so the two work together even when installed separately.
How the project got to 1.0.0
The llama.cpp code history began on March 10, 2023, before Llama 2 came out. In February 2026, the GGML team led by Georgi Gerganov joined Hugging Face. On August 7, 2026, the developers started discussing versioning rules. Next came v0.1.0 and the v0.1.2 pre-release, followed by v0.9.0 on November 4, 2026.
What it means in practice
For AI app developers, version 1.0.0 makes updates more predictable and simplifies adding new models and voice features. For users, it brings ways to use memory more efficiently and additional security settings. These improvements will reach ready-made apps as their developers update them. The release does not guarantee a speedup for every model, though: the team has not published an overall performance comparison with the previous version.
More open-source projects are in our Open source section.
