Autoregressive token generation is notoriously slow and memory-intensive, especially when you only need a model to make a structured classification or routing decision rather than write out paragraphs