llama.cpp Hits 1.0

What happened

llama.cpp and ggml moved to v1.0.0 together on November 13, the first time the project has carried a semantic version number of its own, about three and a half years after it began. Until now the project shipped build-numbered releases such as b12359; from now on, breaking changes to the public C API in either library arrive with a major-version bump, backward-compatible features and new model support with a minor bump, and bug fixes with a patch.

Version 1.0.0 also brings a cleanup: 35 deprecated functions are removed, along with the warmup setter and the defragmentation-threshold option, and the NextN speculative-decoding API moves from the extension header into the main llama.h header. ggml removes deprecated APIs and legacy DIAG_MASK operations and adds configurable YaRN truncation and FP8 support.

New in the release are FP8 key-value caches, a host-memory cache for mixture-of-experts models, and hardening of llama-server against server-side request forgery and DNS rebinding. The project supports more than 150 model-architecture variants; its maintainers joined Hugging Face in February, and the license remains MIT.

Key facts

1. What shipped: llama.cpp and the underlying ggml tensor library both reached v1.0.0 on November 13, the project's first semantically versioned release about three and a half years after it began.
2. Versioning promise: Breaking changes to the public C API now arrive only with a major-version bump; backward-compatible features and new model support come with a minor bump, and bug fixes with a patch.
3. Cleanup: 35 deprecated functions removed, along with the warmup setter and defragmentation-threshold option; the NextN speculative-decoding API moves into the main llama.h header.
4. New features: FP8 key-value caches, a host-memory cache for mixture-of-experts models, and llama-server hardening against server-side request forgery and DNS rebinding.
5. Scale: The project supports more than 150 model-architecture variants; the previous numbered releases were llama.cpp v0.9.0 and ggml v0.29.0.
6. Team: The maintainers joined Hugging Face in February; the project remains under the MIT license.