Many lightweight visualization tools opt to use smaller models (such as GPT-2 or BERT), or pre-compute attention data for offline browsing, in order to balance performance and real-time capability.