Download the model in background. Serve the client with an LLM vendor API just for the first requests, or even using that same local LLM installed on your own servers (likely cheaper). By doing so, in the long run the inference cost is near-zero and allows to use LLMs in otherwise impossible business models (like freemium).