Comment by transitorykris

To expand a bit for my current use cases. Inline routing of work to heavy task specific models, and prompt/context generation (user is asking something, what and how much should we prompt the expensive LLM with). Latency or time to first token does matter for some applications.