Oleksandr Kuzmenko, EditorPublished 6 min readAI-assisted · editor-reviewed
TL;DR
- llama.cpp and ggml adopt strict semantic versioning; breaking C API changes under ./include trigger a major bump.
- Hardware backends can now be built as dynamic libraries and loaded on demand at runtime.
- Supports over 150 architectures and all mainstream speculative decoding techniques out of the box.
Key facts
- Supported Architectures
- 150+ variants
- Time Since Inception
- 3.5 years
- API Boundary
- ./include (public C API)
- Supported Silicon
- NVIDIA, AMD, Qualcomm, Arm, IBM, Intel
Source: Key facts — Supported Architectures, Time Since Inception
Source: Key facts. Supported Architectures: 150+ variants. Time Since Inception: 3.5 years. Each figure is copied from the story.
- Supported Architectures150+ variants
- Time Since Inception3.5 years
Strict Semantic Versioning for Downstream Stability
After 3.5 years of rapid iteration, the ggml-org team has released llama.cpp v1.0.0 and ggml v1.0.0. As adoption has grown from hobbyist scripts to commercial desktop and agentic deployments, maintaining API predictability became crucial. The project now enforces strict semantic versioning: any breaking change to the public C API residing in ./include triggers a major-version bump, while backward-compatible features, new model architectures, and backend additions will land in minor and patch releases.
Dynamic Runtime Backends Across Silicon
The underlying ggml tensor engine has decoupled execution targets into dedicated dynamic backends. Backends can now be compiled as standalone dynamic libraries (.so / .dylib / .dll) and loaded optionally at runtime. This modular approach spans hardware integrations from NVIDIA, AMD, Qualcomm, Arm, IBM, and Intel. The release also ships with expanded CI infrastructure supported by Hugging Face Jobs and Buckets, testing self-hosted runners across all target architectures on every commit.
Maintained Model and Decoding Breadth
The 1.0.0 baseline retains compatibility with more than 150 neural network architectures, covering standard dense transformers, mixture-of-experts (MoE), and vision-language models. It also natively standardises mainstream speculative decoding methods, allowing developers to accelerate inference speeds without custom forks.
Try it in 2 minutes
git clone --branch v1.0.0 https://github.com/ggml-org/llama.cpp
cd llama.cpp && cmake -B build -DGGML_BACKEND_DL=ON
cmake --build build --config ReleaseWhen to use
- Embedding offline or privacy-first inference into desktop and mobile software.
- Running quantised multimodal models on local hardware without Python runtime dependencies.
- Developing cross-platform agents that require fast cold starts and minimal memory overhead.
What to do today
- Pin your project dependency on llama.cpp to tag v1.0.0 in git submodules or CMake fetching scripts.
- Audit C/C++ integrations against headers in ./include to ensure alignment with the new stable baseline.
- Explore compiling platform-specific backends as dynamic libraries to reduce initial binary footprints.
Step names are taken from What to do today.
Step names are taken from What to do today. Pin your project dependency on llama.cpp to tag v1.0.0 in git submodules or CMake fetching scripts.. Audit C/C++ integrations against headers in ./include to ensure alignment with the new stable baseline.. Explore compiling platform-specific backends as dynamic libraries to reduce initial binary footprints.
- Pin your project dependency on llama.cpp to tag v1.0.0 in git submodules or CMake fetching scripts.
- Audit C/C++ integrations against headers in ./include to ensure alignment with the new stable baseline.
- Explore compiling platform-specific backends as dynamic libraries to reduce initial binary footprints.
What the community says
Let's go!!
THIS
