ONNX, the Open Neural Network Exchange, is an open standard and open-source ecosystem for representing machine-learning models in a portable graph format. It defines common operators and model structures so trained models can move between frameworks, runtimes, optimization tools, and hardware backends without being tied to one training stack.

The repository contains the ONNX specification, operator definitions, model schemas, serialization support, validation tools, and reference components used by the broader ecosystem. Frameworks can export models to ONNX, while inference engines and hardware vendors can implement the standard to run those models efficiently.

ONNX is foundational infrastructure for ML interoperability rather than an end-user modeling framework. It is most useful when teams need a stable exchange format between training and deployment systems, especially in environments where models must target multiple runtimes or specialized accelerators.