npr · Kimi K3 run 2
50.5Library score
Pass rate83.6%
Simplicity61.8
Problems $$7.85
Library $$9.66
The library
npr · designed by Kimi K3 (mini-SWE)6,525 source lines
ImplementerPass rateSimplicityCostScore
DeepSeek V4.1 Flash · mini-SWE70.6%56.3$1.70
GLM 5.3 Flash · mini-SWE90.4%69.8$4.61
GPT-5.6 Luna · Codex89.9%59.4$1.53
ProblemDeepSeek V4.1 FlashGLM 5.3 FlashGPT-5.6 LunaMean
cluster-consensus41.714.334.430.1logistic-path72.889.569.877.4relu-regions0.019.126.215.1table-reasoning0.031.186.939.3embedding-pooling59.890.459.669.9kv-cache0.397.556.551.4channel-stft54.675.044.658.1irls-path62.467.356.262.0peak-covariance0.726.622.916.8envelope-curvature59.38.857.341.8radiance-moments50.556.558.455.1kalman-update72.391.978.180.8csi-pose54.765.40.140.1vmc-update0.10.20.10.2pose-correction68.185.069.874.3spectral-reconstruction39.364.357.353.6portfolio-risk79.094.552.075.2contact-forces53.866.659.760.0density-propagation40.741.457.946.6tensor-train0.0100.087.962.6