What's new in TensorFlow 2.21- Google Developers Blog

TensorFlow 2.21 is released, with the LiteRT stack (formerly TFLite) fully graduating into production. LiteRT offers 1.4x faster GPU performance than TFLite, new state-of-the-art NPU acceleration, a unified workflow for GPU/NPU acceleration, superior cross-platform GenAI deployment for models like Gemma, and first-class PyTorch/JAX support. TF.lite operators now support lower-precision data types (int8, int16x8, INT2, INT4) for SQRT, comparison, cast, slice, and fully_connected operators. Community updates focus on security and bug fixes, timely dependency updates, and community contributions for TF.data, TensorFlow Serving, TFX, and other TensorFlow projects. TF Lite is renamed to LiteRT and is in active development separately. Keras 3, JAX, and PyTorch are recommended for new work in Generative AI.