This is a great minimal example! You’re on the right track regarding the input, it’s just that coreml expects the input shape to be fully defined, meaning it must be a 2D tensor of (batch_size, sequence_length).

If you change the conversion inputs line to be

inputs=[ct.TensorType(shape=(32, max_len), name="embedding_input", dtype=np.int32)],

instead of a 1-dimensional tensor you should be fine.

Also you may need to use mlpackage instead of mlmodel for the file extension.

Yep, it essentially just provides a bunch of useful building blocks for networks and some time-saving tools so you don’t have to create them yourself using TensorFlow primitives.

It could be to do with Keras 3 becoming the new default after 2.16.0. They also deprecated tensorflow-macos in favor of a unified installation method so it’s possible internal changes made in the consolidated package account for the recent improvements.