I finally got it working as follows: python -m pip uninstall tensorflow-macos python -m pip uninstall tensorflow-metal conda install -c apple tensorflow-deps --force-reinstall python -m pip install tensorflow-macos python -m pip install tensorflow-metal which gave me the currently latest versions tensorflow-deps 2.10.0 tensorflow-macos 2.11.0 tensorflow-metal 0.7.1 No tensorflow-deps 2.11.0 is available at time of writing. The Apple Developer test script is outdated and can be adjusted to use legacy Adam as follows: import tensorflow as tf cifar = tf.keras.datasets.cifar100 (x_train, y_train), (x_test, y_test) = cifar.load_data() model = tf.keras.applications.ResNet50( include_top=True, weights=None, input_shape=(32, 32, 3), classes=100,) loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True) opt = tf.keras.optimizers.legacy.Adam(learning_rate=0.001) model.compile(loss=loss_fn, optimizer=opt, metrics=["accuracy"]) model.fit(x_train, y_train, epochs=5, batch_size=64) This works now and model.fit was run in appr. 4m10s on M2 Pro Mac Mini.