tfjs

It leverages WebGL for GPU-accelerated computation in browsers that support it, and WebAssembly for efficient CPU execution across all modern browsersachieving inference speeds fast enough for real-time applications like pose estimation, object detection, and language model inference within a web page.

The library provides three usage levels: running pretrained models from the TensorFlow.js model hub directly, converting existing TensorFlow or Keras models to the browser-compatible format using a converter tool, and building and training custom models in JavaScript using the Layers API that mirrors Keras's interface.

Browser-side training is supported for lightweight personalization use cases where adapting a model to individual user data without sending that data to a server is valuableon-device learning without privacy risks.

Web developers building interactive AI demos, educators creating accessible AI demonstrations that run in any browser without installation, developers building privacy-sensitive applications where inference data should never leave the user's device, and creators of real-time AI interactive art use TensorFlow.js as the foundation for browser-native AI capabilities.

The elimination of server infrastructure for inferenceno GPU server to provision, no API to maintain, no per-inference costdramatically lowers the barrier to deploying AI features in web applications and static sites.