BytedTsinghua-SIA/CUDA-Agent
★ 1,344 stars
Covered in week 2026-W35 · the 2026-08 report
CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation
What CUDA-Agent is
CUDA-Agent is the first known RL-trained model to surpass advanced models such as Claude Opus-4.6 and Gemini 3 Pro on high-performance CUDA kernel generation. It achieves state-of-the-art results on KernelBench, consistently outperforming the torch.compile baseline across difficulty levels, with especially strong gains on the hardest cases. To support the LLM-based CUDA generation community, we have released our training data, expert-designed SKILL.md and agent environment.
From the repository README · read it on GitHub →
Transcript
Writing optimized CUDA kernels manually usually means weeks wrestling with GPU memory, thread sync, and performance bottlenecks—only to end up with code slower than hardware limits. CUDA-Agent automates kernel generation using reinforcement learning. The AI agent explores the optimization search space and generates high-performance kernels that rival expert-written code. 1,160 stars, backed by ByteDance and Tsinghua University. Would you trust AI to write your GPU code?