On April 15, 2026, the AI Platform team at Xiaohongshu officially announced the open-sourcing of Relax—a sophisticated training engine tailored for large-scale model reinforcement learning in multi-modal and Agentic applications. Leveraging high-performance backends such as Megatron-LM and SGLang, Relax is capable of unified processing and seamless integration of diverse data types, including images, text, audio, and video. This capability significantly boosts the efficiency and scalability of multimodal training processes. During testing on the Qwen3-Omni-30B model, Relax exhibited consistent and stable convergence in reinforcement learning training across all four modalities. Furthermore, in a multi-machine setup with 16xH800 configurations, fully asynchronous training with Relax achieved a remarkable 76% increase in end-to-end speed when compared to the Colocate baseline, and a 20% improvement over veRL.