TF-Keras Nightly

2.21.0.dev2026032809 verified Tue May 12 auth: no python install: stale

TF-Keras is a deep learning API written in Python, running on top of the machine learning platform TensorFlow. It was developed with a focus on enabling fast experimentation and providing a delightful developer experience. As a nightly build, it offers the very latest features and bug fixes, with daily releases. The current version is 2.21.0.dev2026032809.

pip install tf-keras-nightly

Warnings

breaking Keras optimizers were moved to `tf.keras.optimizers.legacy` starting with TensorFlow 2.12. Directly instantiating `keras.optimizers.Adam()` (or other optimizers) will use the new Keras 3 optimizers, which handle state differently and may break checkpoint loading from models trained with older optimizers. ↓

fix If encountering checkpoint loading issues or needing to use the old optimizer behavior, explicitly use the legacy optimizers: `tf.keras.optimizers.legacy.Adam()`.

breaking The default Keras model saving format for `model.save()` is now the `.keras` format (Keras V3 format), not the legacy H5 format, when using `model.save('my_model.keras')`. This change was introduced with TensorFlow 2.13 and Keras 2.13. ↓

fix To save in the legacy H5 format, explicitly specify `save_format="h5"`: `model.save('my_model.h5', save_format="h5")`.

gotcha `tf-keras-nightly` is a nightly build, meaning it's built from the latest development branch and is not guaranteed to be stable or backward compatible between daily releases. ↓

fix Use `tf-keras-nightly` for access to the newest features and bug fixes, but be prepared for potential instability or API changes without notice. For production environments, prefer stable releases of `tensorflow`.

gotcha Requires Python >=3.10. Installing with older Python versions will result in an installation error. ↓

fix Ensure your Python environment is version 3.10 or newer before installing `tf-keras-nightly`.

breaking `tensorflow` module not found. This typically means the the `tensorflow` package has not been installed in the current Python environment. ↓

fix Ensure `tensorflow` is installed by running `pip install tensorflow` in your environment.

Install compatibility stale last tested: 2026-05-12

python os / libc status wheel install import disk

3.10 alpine (musl) timeout - - - -

3.10 alpine (musl) - - - -

3.10 slim (glibc) timeout - - - -

3.10 slim (glibc) - - - -

3.11 alpine (musl) wheel - - 37.6M

3.11 alpine (musl) - - - -

3.11 slim (glibc) timeout - - - -

3.11 slim (glibc) - - - -

3.12 alpine (musl) wheel - - 29.0M

3.12 alpine (musl) - - - -

3.12 slim (glibc) timeout - - - -

3.12 slim (glibc) - - - -

3.13 alpine (musl) wheel - - 28.6M

3.13 alpine (musl) - - - -

3.13 slim (glibc) timeout - - - -

3.13 slim (glibc) - - 9.81s 2.2G

3.9 alpine (musl) timeout - - - -

3.9 alpine (musl) - - - -

3.9 slim (glibc) timeout - - - -

3.9 slim (glibc) - - 7.55s 2.1G

Imports

keras
```
from tensorflow import keras
```
keras
wrong
```
import keras
```
correct
```
import tensorflow.keras as keras
```
While 'keras' (the multi-backend Keras 3 package) and 'tf.keras' often expose the same API, 'tf-keras-nightly' specifically refers to the Keras implementation bundled with TensorFlow. Directly importing 'keras' might point to a different Keras installation or an older Keras 2 if not explicitly managed.

Quickstart last tested: 2026-04-24

This quickstart demonstrates how to define, compile, train, and save a basic Keras Sequential model using the `tensorflow.keras` API. It uses dummy data for a simple classification task.

import tensorflow as tf
import os

# Ensure Keras backend is set to TensorFlow if running multi-backend Keras 3
# (though tf-keras-nightly implies TensorFlow backend by nature)
# os.environ["KERAS_BACKEND"] = "tensorflow"

# Import Keras from TensorFlow's namespace
from tensorflow import keras
from keras import layers

# Define a simple sequential model
model = keras.Sequential([
    layers.Dense(64, activation='relu', input_shape=(784,)),
    layers.Dense(64, activation='relu'),
    layers.Dense(10, activation='softmax')
])

# Compile the model
model.compile(
    optimizer=keras.optimizers.Adam(learning_rate=0.001),
    loss=keras.losses.SparseCategoricalCrossentropy(),
    metrics=[keras.metrics.SparseCategoricalAccuracy()]
)

# Generate dummy data
import numpy as np
x_train = np.random.rand(100, 784).astype('float32')
y_train = np.random.randint(0, 10, 100).astype('int32')

# Train the model
print("Starting model training...")
model.fit(x_train, y_train, epochs=1, batch_size=32)
print("Model training finished.")

# Save the model (using the recommended .keras format)
model.save('my_model.keras')
print("Model saved as my_model.keras")

# Load the model
loaded_model = keras.models.load_model('my_model.keras')
print("Model loaded successfully.")