r1-lite-preview from DeepSeek surpasses o1-preview in reasoning