Desktop pet programs typically do not contain large language model weights; instead, they call cloud or locally deployed LLM services (such as OpenAI API, Ollama, etc.) via HTTP requests.