Higher token throughput is great for use cases where the smaller, faster model still generates acceptable results. Final response time improvements feel so good in any sort of user interface.