CUDA Agent outcodes Gemini and Opus on auto-written GPU kernels

CUDA Agent is ByteDance’s AI system for automatically writing and optimizing GPU kernels, the low-level code driving deep learning workloads. It writes CUDA code, runs tests, measures performance, and iteratively improves speed, acting as a specialized coding assistant for GPU optimization.

Benchmarks show CUDA Agent beating top generalist models like Gemini 3 Pro and Claude Opus 4.5. This is across general correctness and speed metrics for GPU kernels.

The improvements translate to faster training and inference when integrated into real-world AI pipelines. The project is fully open with the GitHub repo including the full dataset used to train the agent.

Viewers can also check out its detailed workflow diagrams so others can replicate or extend it. For teams building custom CUDA kernels, especially in research or infrastructure companies, CUDA Agent could dramatically reduce development time.