The SGLang RL team has developed an effective INT4 Quantization-Aware Training (QAT) pipeline, inspired by the Kimi K2 team. This approach combines fake quantization during training with real quantization at inference (W4A16), achieving stability and performance similar to BF16 full precision. By compressing models to INT4, large models (~1TB) can now run on a single H200 GPU, eliminating the need for cross-node communication and boosting rollout efficiency. The implementation involves a closed-loop system where BF16 weights are used during training, and quantization noise is simulated through fake quantization. This method forces the model to adapt to lower precision while maintaining high precision in calculations. The training uses a Straight-Through Estimator (STE) to allow gradients to flow through the quantization layers, updating the BF16 weights effectively. The final step converts these weights into INT4 for inference, allowing efficient execution with W4A16. Key benefits of this system include reduced VRAM and bandwidth pressure for large models, achieving better rollout efficiency than previous methods like W8A8. The team also provides a detailed technical recipe and welcomes contributions from the community. Future plans include exploring FP4 quantization on NVIDIA Blackwell GPUs, further pushing the boundaries of low-precision training. Overall, this project represents an important step in enhancing the efficiency and scalability of reinforcement learning models.