In May, KDA placed in the top three on every track of the MLSys 2026 FlashInfer kernel contest. That was a good result against the field. It was not a good result against the field's best entries, which were written by people.
The second generation of KDA changes three things and reruns the contest.
Swipe the diagram sideways to see all of it →
1/0 The loop: an agent writes a kernel, a verifier measures and reviews it.
- CuteDSL support, so the agent can express the layouts it wants rather than the ones the previous toolchain made convenient;
- the IKET profiling skill, which turns a profile into something an agent can reason from instead of a wall of counters; and
- a better agent loop flow: the arrangement of who writes, who checks and what carries between attempts.
The rerun
The kernels are now public, with a benchmark that pins the human winners' own repositories (Kachua on GDN, Team Wombat on MoE, Dogacel on DSA) and refuses to run if they are modified. The chart opens against the best human entry; switch the baseline to the FlashInfer baseline or the contest entry, or switch to the B300 numbers this post first reported.
The MoE row is the one worth sitting with. The contest entry was slower than the baseline there, at 0.67×, and the same application, with a profiling skill and a different loop, is now past the best human entry. Nothing about the model changed.The public repository numbers the two generations KDA 0.1 and KDA 0.5; this post first called them KDA 1.0 and 1.5.
GDN prefill · over Kachua
1.688×The largest margin over a human winner, at 10.36× the FlashInfer baseline.
MoE FP8 · contest entry → now
1.57×From slower than the baseline to 1.17× the best human entry.
The benchmark's own caveats apply. Tuning targets the official shapes, so do not expect the same speedups elsewhere, and the pinned human repositories are a few commits past the contest deadline.
A workload nobody had tuned
Contest tracks have been optimised by many people, which makes them a fair test and a stale one. So we also pointed the new loop at k-means for video diffusion, which nobody had worked over.
Update, October
That work became a CuTe backend for Flash-KMeans, measured at 7.31× on a real LongCat-Video workload on B200. It was merged on 27 August and reverted the next day; the October roundup has the details. An earlier version of this post gave 6.1× on a Wan 2.2 workload, a figure with no public source, so it is gone.