llama.cpp v1.0.0: Stable Release with New Architectures and TTS

llama.cpp v1.0.0 has been released. This is the project's first stable release: deprecated functions have been removed from the public API, the NextN API has been moved into the main llama.h header, and ggml has been updated to version 1.0.0. Notable additions include FP8 KV caches, a MoE expert cache in system RAM with a configurable CPU share, loading of external PLE shards, three new text architectures (Kimi-K3.1-Pro, Ripple, xing4_0), the drama-3 TTS model, vision support for MageVL-Cowork and Kimi-K3.1, and the /v1/audio/speech endpoint in llama-server.

What it means

For those running models locally, the release brings several practical changes. First, the update to v1.0.0 means a break with the deprecated API: code written against the old llama.h functions will need edits. Second, the MoE expert cache in system RAM lets you offload VRAM to the CPU — the developer provides a way to configure what share of experts stays on the processor. FP8 KV caches reduce memory consumption when working with long contexts. The expanded list of supported architectures means the Kimi-K3.1-Pro, Ripple-65B-Preview and xing4_0 models can now be run through llama.cpp without additional layers. The arrival of drama-3 TTS and the /v1/audio/speech endpoint adds speech generation to llama-server, while the updated web interface with chat groups and a markdown editor changes how you work with the server. The developer does not provide exact memory requirements or parameters for the new models in the release description. You can follow updates on tools for running neural networks locally in the CPU3D news feed.