how does deepseek r1's reasoning capability compare to perplexity's