GPU kernel Very high Verified
Cursor NVIDIA kernel optimizations
Kernels produced by Cursor's multi-agent system in a collaboration with NVIDIA, reported to reach a 38% geomean speedup on 235 Blackwell problems.
The 14 April 2026 Cursor blog describes a planner-and-worker multi-agent run over three weeks, writing CUDA C with inline PTX and CuTe DSL, benchmarked with SOL-ExecBench on Blackwell 200 GPUs. Final artifacts are published in anysphere/kernel-optimization-results.
Human involvement
Mostly written by agents. The blog presents the run as autonomous after the experiment was set up, with a planner distributing work across workers. Model names for that run were not extracted from the post.
Evidence
- Official blog
Our multi-agent system autonomously optimized 235 CUDA kernels for NVIDIA Blackwell 200 GPUs, achieving a 38% geomean speedup over baselines in just 3 weeks.
- Official repository