ONNX

Open standard for representing and exchanging machine learning models across frameworks and runtimes.

Free· Free and open source (Apache-2.0); Linux Foundation AI projectFine-tuning

Best for

Pick ONNX if you need to move a trained model out of its source framework and into a different runtime, hardware target, or edge device.

Skip if

Skip it if you're staying inside a single framework end-to-end or want a managed inference platform rather than a file format.

ONNX (Open Neural Network Exchange) is an open format for representing machine learning models, defining a shared set of operators and a common file structure that any compliant framework, runtime, or compiler can read. In practice it lets you train a model in PyTorch, export it to ONNX, then run it through ONNX Runtime, TensorRT, OpenVINO, CoreML, or a browser-side engine without rewriting the model.

It's not a SaaS product and there's no pricing - ONNX is a Linux Foundation AI graduate project maintained by a multi-vendor community (Microsoft, Meta, NVIDIA, Intel, AMD, and others). The audience is ML engineers and platform teams who need framework portability, hardware acceleration across CPUs/GPUs/NPUs, or edge deployment where the training stack isn't available at inference time. The actual runtime work happens in adjacent projects like ONNX Runtime; ONNX itself owns the spec, the opset, and the reference tooling.

The ecosystem is broad: exporters in PyTorch and TensorFlow/Keras, runtimes from most major hardware vendors, optimization tools (onnx-simplifier, onnx-optimizer), and quantization paths for INT8/FP16 deployment. Caveats are real - opset version mismatches and unsupported ops still cause export failures, and dynamic-graph or custom-op models can need manual surgery before they cleanly serialize.

Editor's take

ONNX is plumbing, not a product, but it's the plumbing the entire model-deployment industry has standardized on. If you're shipping inference anywhere outside the training framework, you'll touch it eventually. Expect to spend time on opset compatibility - that's the cost of portability.

— The AI Tool Bible editorial team

Pros

✅ Vendor-neutral standard backed by Linux Foundation and every major hardware maker
✅ Export once, deploy to CPUs, GPUs, NPUs, mobile, and browsers via compatible runtimes
✅ Mature tooling for quantization, graph optimization, and opset conversion
✅ Massive ecosystem of pretrained models available in ONNX format