ONNX is an open format for AI models, covering both deep learning and traditional machine learning. It defines an extensible computation graph model, built-in operator definitions and standard data types, so a model trained in one framework can run in other tools and on other hardware; the current focus is inference. The repository dates to September 2017, the project is listed by LF AI & Data, and the latest release is 1.23.2 (6 October 2026). Apache 2.0.
ONNX
ONNX is the interchange format for trained models. It describes a model as a computation graph of nodes that call standard operators over typed tensors, so the same file can move from a training framework to a runtime, compiler or accelerator that never saw the original code.
- Intermediate representation - The IR spec defines
the extensible graph model and its data types; the
.protofiles form the normative syntax, with a model represented asonnx.ModelProto. - Operator sets - Built-in operators are documented per version, and new operators go through a documented proposal process.
- Versioned independently - The IR, the operator sets and the library each carry versions, described in the project’s versioning principles, with an opset version converter for moving between them.
- Tooling - The project documents shape and type inference, graph optimization and opset version conversion, and links pre-trained ONNX models on Hugging Face.
- Open governance - A Steering Committee, working groups and Special Interest Groups meet on a public calendar. Apache 2.0.
Where it sits. ONNX is to model weights what OpenAPI is to HTTP interfaces: a neutral description that lets producers and consumers come from different vendors. Its serialization builds on Protocol Buffers, and it sits underneath the model-serving APIs that MLOps practice wraps around inference.
