Saving is for you, exporting is for the world
Everything so far has been about save — round-tripping a model so you can keep training it. export is the opposite direction: it produces an inference-only artifact that a serving system can run, often one that has never heard of Keras. The single model.export() call takes a format= argument, and that one argument picks which deployment world you're targeting.
One call, four destinations
Each format answers a different deployment question. tf_saved_model is the TensorFlow ecosystem's native serving format — feed it to TF Serving or TF.js. onnx is the vendor-neutral interchange standard; export to ONNX when the inference runtime (Triton, ONNX Runtime, a hardware vendor's stack) expects it. openvino targets Intel's optimized inference engine for CPU-bound deployments. litert (the format formerly branded TFLite) is the mobile and edge target — small, quantizable, built to run on phones and microcontrollers.
The point isn't to memorize four strings. It's that the model you trained is a single source of truth, and exporting is a projection of it onto whatever runtime your deployment happens to need.