ONNX is an open format for representing machine learning models, developed initially by Microsoft and Facebook in 2017. The purpose is interoperability — a model trained in PyTorch can be exported to ONNX and then run in TensorRT, ONNX Runtime, or any other compliant inference engine, without being rewritten.
In enterprise ML deployment, ONNX is the canonical intermediate format for moving trained models from the research environment into production inference systems. It matters because training and inference have very different hardware and performance requirements, and separating the two via an interchange format lets teams optimize each independently.
ONNX's adoption has been uneven. Frameworks sometimes lag in supporting new operators, and the performance characteristics of ONNX-exported models vary. For large language models specifically, most deployment today uses model-specific runtimes (vLLM, TensorRT- LLM) rather than ONNX. But for the broader universe of computer vision, audio, and classical ML models, ONNX remains the dominant interoperability standard.