TopolearnYou need to resume training, share models and deploy them; each needs a different save format.
Keras 3 has three ways to persist a model:
- `model.save("model.keras")`: the native format, a zip with the architecture config, weights and optimizer state. Reload with
keras.saving.load_model("model.keras"); the result is ready topredictor continuefit. Custom layers must be registered or passed viacustom_objects. - `model.save_weights("model.weights.h5")` and
load_weights(...): weights only. The file name must end in.weights.h5, and you rebuild the same architecture in code before loading. - `model.export("dir")`: an inference-only artifact. By default it writes a TensorFlow SavedModel (needs TensorFlow installed) for TF Serving or LiteRT conversion; recent versions also offer
format="onnx".
The legacy .h5 full-model format and the old SavedModel mode of model.save() from Keras 2 are no longer the way to save in Keras 3.
The .keras format is backend-independent: a model trained with JAX can be loaded under PyTorch. For checkpoints during training, use the ModelCheckpoint callback, which writes the same formats.
Loading runs in safe_mode by default, which refuses to deserialise arbitrary Python code such as Lambda layers from untrusted files.
Example
import numpy as np
import keras
model = keras.Sequential([keras.Input((3,)), keras.layers.Dense(2)])
model.compile(optimizer="adam", loss="mse")
model.save("model.keras") # config + weights + optimizer
restored = keras.saving.load_model("model.keras")
x = np.ones((1, 3), dtype="float32")
print(np.allclose(model.predict(x, verbose=0),
restored.predict(x, verbose=0))) # True
model.save_weights("model.weights.h5") # weights only
clone = keras.Sequential([keras.Input((3,)), keras.layers.Dense(2)])
clone.load_weights("model.weights.h5") # same architecture required
print(clone.get_weights()[0].shape) # (3, 2)
# For serving (requires TensorFlow): model.export("export_dir")Remember
.keras= full model, backend-independent.weights.h5= weights only, rebuild architecture firstmodel.export()produces inference artifacts for serving
Watch out for
- Saving weights to a file not ending in
.weights.h5, which Keras 3 rejects.