Comparing Transformers and Hybrid Models at the Token Level