After 3.5 years of development, the Hugging Face team behind llama.cpp has tagged v1.0.0 for both libllama and the tensor library under it, ggml v1.0.0. There is no new feature to announce here: the milestone is a contract, and the contract is the point. Every downstream binding, desktop app and packaging pipeline builds against the two public C headers in include/, and until now a breaking change could land in any release.

From 1.0 the semantic version rules are explicit:

  • MAJOR: a breaking change to the public C API in ./include
  • MINOR: a backward-compatible feature, backend update, model support or API addition
  • PATCH: a bug fix with no API change

API and ABI incompatibilities now show up as major version bumps, so a consumer can pin a version line instead of chasing master.

Distribution is already broad, which is why the guarantee matters: pre-release artifacts for Linux, macOS, Windows, Android and iOS, plus Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework for iOS and macOS. The library covers more than 150 model architecture variants and all mainstream speculative decoding methods, with ggml backends compiled as dynamic libraries that can be loaded at runtime.

The release post also details the testing setup behind that promise: CI runs on self-hosted runners spanning NVIDIA, AMD, Qualcomm, Arm, IBM and Intel hardware, scaled with HF Buckets and HF Jobs, on the theory that a stable ABI is worth only as much as the coverage that catches an accidental break.

Related: PrismML Releases Bonsai 27B, Extreme Compression for Local AI, Qwen3.8-Flash Now Run Locally via Unsloth GGUF

llama.cpp v1.0.0 announcement · llama.cpp release · ggml release