Xiaohongshu's AI team has open-sourced Relax, a reinforcement learning engine for multimodal training that efficiently processes image, audio, and video inputs through advanced parallelism and asynchronous pipelining.