DeepSeek released its V4.1-Flash model, a 552-billion-parameter, multimodal, mixture-of-experts model with a 1-million-token context window