npr · Kimi K3 run 1
54.4Library score
Pass rate89.5%
Simplicity62.8
Problems $$6.85
Library $$10.87
The library
npr · designed by Kimi K3 (mini-SWE)6,513 source lines
ImplementerPass rateSimplicityCostScore
DeepSeek V4.1 Flash · mini-SWE92.6%63.7$1.78
GLM 5.3 Flash · mini-SWE90.1%68.3$3.51
GPT-5.6 Luna · Codex85.8%56.5$1.56
ProblemDeepSeek V4.1 FlashGLM 5.3 FlashGPT-5.6 LunaMean
cluster-consensus15.725.733.625.0logistic-path95.487.70.061.0relu-regions41.315.716.324.4table-reasoning91.582.678.384.1embedding-pooling69.679.959.769.7kv-cache67.658.160.962.2channel-stft58.766.851.058.8irls-path46.570.462.459.8peak-covariance1.145.733.026.6envelope-curvature40.446.347.744.8radiance-moments60.660.860.960.8kalman-update77.895.266.579.9csi-pose63.061.40.041.5vmc-update65.30.264.543.3pose-correction64.689.45.053.0spectral-reconstruction60.551.840.651.0portfolio-risk25.676.567.756.6contact-forces61.463.757.260.8density-propagation55.83.956.138.6tensor-train85.2100.070.785.3