What you'll reach for, and when
TensorFlow proper (the core library) is the entry point. The ecosystem around it is bigger than the core. Knowing the map prevents you from re-implementing things polished for years.
| Library | Purpose | 2026 status |
|---|---|---|
tf.keras | High-level model API (Keras 3 underneath since TF 2.16) | Primary API |
tf.data | Data pipelines | Core — won't go anywhere |
| TensorBoard | Training visualization, profiling | Active (separate pip install from TF 2.21+) |
| TFLite → LiteRT | Mobile + edge inference | Migrating to LiteRT |
| TF Serving | Production REST/gRPC server | Active |
| TF.js | Browser + Node inference | Active |
| KerasCV | Detection, segmentation, augmentation | Active |
| KerasHub (was KerasNLP) | Pretrained LLMs (Gemma, LLaMA 3, BERT, GPT-2) | Active — replacing KerasNLP |
| TF Probability | Bayesian inference, MCMC, normalizing flows | Active |
| TF Decision Forests / YDF | Random forests, gradient-boosted trees | Active — strong on tabular |
| TF Recommenders | Retrieval + ranking systems | Active |
| TF Agents | RL: DQN, PPO, SAC, TD3 | Active |
| TF Federated | Federated learning on decentralized data | Active research |
The unifying theme: everything in this table speaks Keras. A KerasCV YOLOv8 detector, a KerasHub BERT classifier, and a TF-DF gradient-boosted tree all expose .fit, .predict, and .save. Once you know the Keras shape, the rest of the ecosystem is one pip install away.