llama.cpp reaches 1.0 with changes for local AI developers

llama.cpp has released version 1.0.0, a major update to the open-source inference project used to run language models on local hardware. The November 13 release notes describe changes to developer APIs, model support, the server and its web interface.

Why the major version matters

The maintainers removed deprecated public interfaces and promoted speculative decoding interfaces into the main header. That is relevant to applications that integrate llama.cpp as a library: downstream developers should inspect the release notes and test their integrations before upgrading. A version number is not a guarantee that a particular launcher or plugin will work unchanged.

The update also lists a configurable host-memory cache for mixture-of-experts models, additional vision and text architectures, and support for an OpenAI-compatible audio speech endpoint. The project describes changes to server protections involving server-side request forgery, cross-origin requests and DNS rebinding. Read the release notes and deployment documentation for the exact scope; an update does not replace a security review.

What local users should expect

llama.cpp is an inference engine, not a model. Supported hardware, available memory, the model format and quantization still determine what can run. Features added to the engine do not automatically appear in every application built on top of it.

For developers packaging local tools, the important work is compatibility testing: back up working configurations, review removed APIs and verify model loading, streaming and cancellation with the actual frontend. We have not installed or benchmarked this release.