DeepSeek claims its R1 model outperforms rival models in mathematical tasks, general knowledge, and question-and-answer performance.