OpenRouter-served models are often more heavily quantized than what you can run locally, or try for yourself on generic cloud-based infrastructure.