[{"data":1,"prerenderedAt":12},["ShallowReactive",2],{"stack-tr-en-onnx-runtime":3},{"tagline":4,"description":5,"strengths":6,"useCase":11},"The portable inference engine that runs one ONNX model from server to browser.","ONNX Runtime, built by Microsoft, loads a model in the ONNX format exported from PyTorch, TensorFlow or scikit-learn and runs it through interchangeable execution providers: CPU, CUDA, TensorRT, OpenVINO, DirectML, CoreML, NNAPI, or WebAssembly and WebGPU in the browser. The graph is optimized at load time (constant folding, operator fusion) and can be quantized to INT8, with APIs for Python, C++, C#, Java and JavaScript. It's our choice when the same model has to run on heterogeneous targets without being recompiled for each one.",[7,8,9,10],"CPU, CUDA, TensorRT, OpenVINO, DirectML and CoreML execution providers: one model, many accelerators.","Load-time graph optimizations plus static or dynamic INT8 quantization.","onnxruntime-web: in-browser inference through WebAssembly or WebGPU, no backend needed.","MIT license, Python, C++, C#, Java and JavaScript APIs, GenAI extension for LLMs.","Inference on CPU servers, Windows or macOS desktop apps, and models embedded in the browser.",1792675979950]