ONNX

ONNX (Open Neural Network Exchange) is an open format built to represent machine learning models, enabling models trained in one framework to be transferred to another for inference. Co-created by Microsoft and Facebook (Meta) in 2017, with later contributions from AWS, IBM, and NVIDIA, ONNX has become the industry standard for model interoperability with over 18,200 stars as of 2026. The ONNX format defines: an extensible computation graph model (representing the model as a directed acyclic graph of operator nodes, each with inputs, outputs, and attributes), a standard set of operators (over 170 operators covering tensor operations: Conv, Pool, Gemm, MatMul, BatchNorm, Softmax, ReduceMean, Concat, Split, Reshape, Transpose, Squeeze, Unsqueeze, Gather, Scatter, activation functions: Relu, Sigmoid, Tanh, LeakyRelu, Elu, PRelu; and model metadata: IR version, opset imports, producer info). ONNX defines two model representation levels: the base ONNX format (for inference) and ONNX Training (for representing training graphs with gradient operations). The ONNX Runtime (developed by Microsoft) provides high-performance inference across platforms (Windows, Linux, macOS, Android, iOS, web via WebAssembly) with hardware acceleration via CUDA, TensorRT, OpenVINO (Intel), DirectML (Windows), CoreML (macOS/iOS), NNAPI (Android), ROCm (AMD), and custom execution providers. ONNX supports models from PyTorch (via torch.onnx.export), TensorFlow (via tf2onnx), scikit-learn (via skl2onnx), Keras, XGBoost, LightGBM, and Spark MLlib. The format is governed by the LF AI & Data Foundation.