Announcing ONNX Runtime 1.0 - Open Source Blog

One year after ONNX Runtime's initial preview release, we're excited to announce v1.0 of the high-performance machine learning model inferencing engine. This release marks our commitment to API stability for the cross-platform, multi-language APIs, and introduces a breadth of performance optimizations, broad operator coverage, and pluggable accelerators to take advantage of new and exciting hardware developments. In its first year, ONNX Runtime was shipped to production for more than 60 models at Microsoft, with adoption from a range of consumer and enterprise products, including Office, Bing, Cognitive Services, Windows, Skype, Ads, and others. These models span from speech to image to text (including state of the art models such as BERT) and ONNX Runtime has improved the performance of these models by an average of 2.5x over previous inferencing solutions. In addition to performance gains, the interoperable ONNX model format has also provided increased infrastructure flexibility, allowing teams to use a common runtime to scalably deploy a breadth of models to a range of hardware.