A model trained inside a real GPU development environment now writes CUDA kernels that regularly beat the compiler, according to a paper published Monday by ByteDance Seed and Tsinghua AIR. The system goes by CUDA Agent.
The gap it attacks is narrow: frontier models generate correct CUDA but often slow CUDA. Seed1.6, the base model, clears 74% of KernelBench tasks yet only beats torch.compile on 27.2% of them; its kernels average a 0.69x geometric-mean speedup.
CUDA Agent’s answer is to drop the model into a real development environment with profiling, correctness checks, and a locked-down sandbox, then run PPO for 150 steps at a 131,072-token context. The trained system passes 98.8% of the 250-task benchmark and beats the compiler 96.8% of the time, with a 2.11x geometric-mean speedup, landing roughly 40 points ahead of Claude Opus 4.5 and Gemini 3 Pro on the hardest split.
The trained agent itself is not being released, and the paper ships no weights. Seed1.6 stays proprietary, a mixture-of-experts model with 23B active parameters out of 230B total. The release package instead includes a 6,000-sample dataset called CUDA-Agent-Ops-6K, a SKILL.md specification, and reward and warm-up recipes for teams that want to retrain.
Replication is a big-machine affair, since the profiling sandbox alone consumed 128 NVIDIA H20 GPUs. Teams without that firepower can still build on the released dataset, reward constraints, and skill spec using an open base model.
