ONNX (Open Neural Network Exchange) is the open standard format for representing machine learning models, stewarded by a broad industry community. The project has more than 21,600 stars on GitHub and is actively maintained, with updates in October 2026 under the Apache-2.0 license.
ONNX solves one of machine learning’s most practical problems: models trained in one framework used to be painful to run anywhere else. By defining a shared graph format and operator set, ONNX lets teams train in PyTorch, TensorFlow, Keras, or scikit-learn, export to a single portable file, and then run that model with high-performance runtimes on servers, desktops, mobile devices, and specialized accelerators. It is governed as a vendor-neutral open project, which is why virtually every major framework and hardware vendor supports it.
In day-to-day practice, ONNX sits quietly underneath a huge share of production AI. Recommendation systems, image classifiers, speech models, and increasingly large language models are exported to ONNX for serving because runtimes can fuse operations, apply quantization, and target GPUs, CPUs, and neural accelerators from one artifact. The specification evolves through community working groups and versioned operator sets, so older models keep running while new operators land for modern architectures. For teams, that stability matters: an ONNX model produced today is a portable asset rather than a framework-specific checkpoint that only one stack can open.
ONNX is completely free and open source under the Apache-2.0 license. The specification, reference implementation, and tooling cost nothing to use in commercial or research projects.
ONNX is essential infrastructure for ML engineers who need to move models between training frameworks and production environments, deploy to edge or embedded hardware, or avoid locking their model portfolio to a single vendor stack.