TensorFlow Serving supports a wide range of models for serving purposes. It is primarily designed to serve TensorFlow models, but it also supports other model formats, such as TensorFlow Lite and SavedModel.

  1. TensorFlow models: TensorFlow Serving works seamlessly with models that have been built using the TensorFlow framework. These models can be trained using TensorFlow's high-level APIs like Keras or low-level APIs like tf.function.

  2. TensorFlow Lite models: TensorFlow Lite is a lightweight version of TensorFlow specifically designed for mobile and embedded devices. TensorFlow Serving can serve TensorFlow Lite models for deployment on resource-constrained devices.

  3. SavedModel: TensorFlow's SavedModel format is a language-neutral serialization format for TensorFlow models. TensorFlow Serving can load and serve models in this format, which provides a standard way to package and deploy models consistently across different platforms and languages.

  4. Custom model types: TensorFlow Serving also allows serving custom-built models, as long as they are compatible with TensorFlow's model format. This gives users the flexibility to deploy models created using different frameworks, such as PyTorch or Caffe, by converting them into TensorFlow-compatible formats.

  It is important to note that TensorFlow Serving focuses on serving inference requests rather than model training or retraining. Therefore, it is primarily used for deploying pre-trained models for real-time predictions in production environments.