Microsoft Foundry Generative AI Chat App Practice Question 1: Model Leaderboard Selection

The Foundry model leaderboards benchmark models across quality, safety, cost, and performance. Performance benchmarking exists precisely "to understand how models perform in terms of latency and throughput," and the documented guidance is that "a model with high quality but high latency might not suit real-time applications." Time to first token (TTFT) is the time until the first token comes back when streaming is enabled, which is exactly what a user perceives as the assistant "starting to reply," and generated tokens per second (GTPS) describes how quickly the rest of the answer streams out.

- A is wrong because the quality index averages accuracy scores across reasoning, knowledge, math, and coding datasets. It says nothing about response time, and the docs explicitly warn that high quality can coexist with high latency. - C is wrong because the safety leaderboard reports attack success rates and toxicity detection scores. Those matter for a customer-facing bot, but they don't separate the three candidates on responsiveness. - D is wrong because cost benchmarks report USD per benchmark run. The stem already rules cost out as a constraint.

Reference: Model benchmarks and leaderboards in Microsoft Foundry