Skip to content
← All news

News1 min read

#1 on SOL-ExecBench L1

KDA-generated kernels put the Databricks entry first on SOL-ExecBench's L1 single-operation track, at a SOL score of 0.7639. It is still first on that board, by 0.00006.

The L1 single-operation track of SOL-ExecBench has a new leader: the Databricks entry, with kernels generated by KDA. The first-place score was confirmed on 2 July at 0.7608 (Humanfia-reported), and a submission five days later raised it to 0.7639, which is where the board has it today.

L1 · single operations · 94 kernels · board v1.0
  1. 1Databricks · KDA7 July0.76394
  2. 2ac4k5 July0.76387
  3. 3doubleAI29 June0.75873
  4. 4Recursive9 June0.75843
  5. 5SF Tensor9 July0.72258
  6. 6Argus6 July0.65056
Each entrant's best submission, read 5 October 2026. Hover a row for its gap to the leader.

A SOL score puts each kernel between the scoring baseline (0.5) and the hardware's speed-of-light bound (1.0), and averages over the track. Drag the meter to see where the reference implementation, the baseline and the leaders sit.

Where 0.7639 sits on the SOL scale
  1. 10.237 the reference implementation
  2. 20.5 the scoring baseline
  3. 30.7587 doubleAI, the best score before July
  4. 40.7639 Databricks · KDA
  5. 51.0 speed of light
This score isahead of every earlier entry

A small margin, on a track everyone had tuned ​

The margin is small, and worth being plain about. Before July, the scores to beat were Recursive's 0.7584 (9 June) and doubleAI's 0.7587 (29 June).An earlier version of this post named Recursive's 0.7584 as the previous best. On the board, doubleAI's 0.7587 is a few days older than our result and slightly higher. Today ac4k sits 0.00006 behind. This is not a result where an agent loop found something nobody had thought of. It is a result where a track many people had already worked over moved again, by a little, because the search ran longer and more carefully than a person would sit through.

Over doubleAI, the best before July

+0.0052

0.7587 → 0.7639 on the same 94 kernels.

Over ac4k, second today

+0.00006

A lead that could change with one submission.

That is most of what these applications are for. The dramatic wins make better posts; the narrow ones on picked-over tracks are the better evidence that the method is doing something repeatable.

Since then ​

In September the board moved to a new evaluation stack, v1.1, and rescored the field. On it, the same entry's best L1 score is 0.7963, third, behind Infinigence AI and SF Tensor; across all four collections it holds the most per-kernel first places.

L1 · single operations · board v1.1 · 5 October 2026
  1. 1Infinigence AI2 October0.8032
  2. 2SF Tensor5 October0.7988
  3. 3Databricks · KDA28 September0.7963
  4. 4RIAC7 September0.7886
  5. 5Geometric5 October0.7830

KDA · KDA-Pilot · the board

Read next

News6 KDA PRs merged into SGLangSince August, SGLang has merged six pull requests carrying kernels our agents wrote, the best adding 8.7% end-to-end throughput. The MLSys contest kernels are public with a benchmark anyone can rerun. One merge elsewhere, Flash-KMeans, was reverted the next day.Read it BlogKDA²: Kernel Design Agents (KDA) Optimize Kimi Delta Attention (KDA)Our agents wrote Kimi Delta Attention kernels that run up to 2.96× faster than FlashKDA on B300 with a tenth of its state error. Here is how, and how the agents tried to cheat along the way.Read it News6.5× faster MSA prefill indexerOn B300, KDA found the hardware the MSA prefill and decode indexers were leaving idle: a 6.5× geometric mean on prefill, up to 14× on chunk tails, up to 3.3× on decode, and bitwise-identical output throughout. Humanfia-reported.Read it