GitHub Repo
llama.cpp Releases 1.0, Clearly Marking API Compatibility and Removing Legacy Interfaces
llama.cpp released v1.0.0 on November 13, aligning with ggml to use major version numbers for API and ABI breaking changes. The release also adds an FP8 KV cache and a main-memory MoE expert cache; downstream integrations should first check for removed interfaces.

The llama.cpp team released v1.0.0 on November 13, and the core tensor library ggml also moved to version 1.0. For developers embedding local inference in desktop applications, services, or language bindings, the most important change is the compatibility policy: going forward, API and ABI breaking changes will be reflected in the major version number, making upgrade boundaries easier for downstream projects to identify. Official announcement
The announcement identifies the C API in the public include directory as the basis for versioning: breaking changes increment the major version; backward-compatible features, backends, and model support increment the minor version; and bug fixes that do not change the API increment the patch version. This gives package maintainers a clearer basis for scheduling dependency updates, but it does not mean that all internal interfaces receive the same guarantees. Versioning policy
This upgrade itself entails migration work. The release notes list the removal of llama_set_warmup, 35 deprecated functions, and the defrag_thold field and command-line option. The NextN/speculative decoding interfaces have also moved from an extension header into the public llama.h, with one function renamed. Projects that link directly to the library or maintain their own bindings should compare the headers with their call sites, then rebuild and verify. Release notes
Memory management also sees substantive updates: an OCP FP8 KV cache is available via -ctk/-ctv f8_e4m3, with support for static per-tensor scaling; the MoE expert cache resides in main memory and offers configurable CPU offload ratios and eviction policies. In practice, the former merits long-context tests for capacity and quality, while the latter calls for monitoring cache hit rates and work distribution between CPU and GPU. The feature names alone are not enough to predict the speedup. Technical changes
For deployment teams, the next step is to include version pinning, binding rebuilds, and regression tests on fixed workloads in the update process. Official materials do not yet provide a single performance gain that applies across hardware and models. The main value of 1.0 is clearer compatibility signals for a rapidly evolving inference stack; latency, memory use, and output quality still need to be measured separately.