Open Neural Network Exchange: an open standard for representing ML models that enables portable inference across frameworks and hardware. By 2026, ONNX Runtime
ONNX and ONNX Runtime are the interoperability layer of model deployment: export from any framework, run optimized anywhere, cloud, browser, mobile, edge. For teams shipping models across heterogeneous targets, ONNX is how 'train once, deploy everywhere' actually works.
Pricing: Free and open-source (MIT/Apache).
When deployment targets differ from your training stack, CPU fleets, browsers, mobile, mixed hardware, or you want runtime-level optimization without framework baggage in production.
For small/medium models and edge LLMs, yes; frontier-scale serving typically uses specialized engines. ONNX shines brightest on vision, speech, and classic DL deployment breadth.