ByteDance and Tsinghua present CUDA Agent for faster GPU kernels
ByteDance Seed and Tsinghua AIR have introduced CUDA Agent, a reinforcement-learning system designed to generate CUDA kernels that outperform compiler-generated code. On the 250-task KernelBench benchmark, the underlying Seed1.6 model passed 74.0% of tasks but produced kernels faster than torch.compile in only 27.2% of cases, with a 0.69× geometric-mean speedup.
CUDA Agent places the model in a sandbox with profiling, correctness checks and restricted permissions, then trains it with PPO for 150 steps using a 131,072-token context. Table 1 reports a 98.8% pass rate, a 96.8% rate of beating torch.compile and a 2.11× geometric-mean speedup over it. On the most difficult Level-3 split, the system reached a 90.0% faster-than-compile rate, compared with 50.0% for Claude Opus 4.5 and 52.0% for Gemini 3 Pro.
The trained agent and model weights have not been released. CUDA Agent is based on proprietary Seed1.6, with 23B active and 230B total parameters. Public materials include the CUDA-Agent-Ops-6K dataset, the SKILL.md specification and reward recipes. The profiling sandbox used 128 NVIDIA H20 GPUs.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.