"We train small models when we need to do it. We can do it on on-premise solutions, but we can also use state-of-the-art LLMs."