ByteDance and Tsinghua Release CUDA Agent, an AI System That Writes Faster GPU Code Than Standard Compilers

ByteDance Seed and Tsinghua AIR have developed CUDA Agent, a reinforcement learning system that trains AI models to generate GPU kernels optimized for speed. While frontier language models already produce correct CUDA code, their output often runs slower than compiler-generated code. The system addresses this gap by placing the model in a real development environment with profiling and sandbox protections, then training it using reinforcement learning. The results are striking: the trained model achieves a 98.8% success rate and outperforms torch.compile on 96.8% of test tasks, with speedups averaging 2.11 times faster than traditional compilers. The system also significantly outpaced comparable AI models like Claude Opus 4.5 and Gemini 3 Pro on the most challenging benchmarks. While the trained model remains proprietary, ByteDance released supporting resources including a 6,000-sample dataset, technical specifications and reward training guidelines. The approach requires substantial computational resources, with experiments running on 128 NVIDIA H20 GPUs, making full reproduction feasible primarily for large research labs and infrastructure teams. The technology targets performance-critical applications including AI inference, autonomous driving, medical imaging and recommendation systems.