npr · Kimi K3 run 3
50.8Library score
Pass rate87.9%
Simplicity60.5
Problems $$8.20
Library $$7.12
The library
npr · designed by Kimi K3 (mini-SWE)5,838 source lines
ImplementerPass rateSimplicityCostScore
DeepSeek V4.1 Flash · mini-SWE89.2%60.5$1.73
GLM 5.3 Flash · mini-SWE85.9%59.6$4.88
GPT-5.6 Luna · Codex88.5%61.3$1.58
ProblemDeepSeek V4.1 FlashGLM 5.3 FlashGPT-5.6 LunaMean
cluster-consensus32.316.534.727.8logistic-path100.059.0100.086.3relu-regions16.513.032.020.5table-reasoning62.723.673.153.1embedding-pooling65.782.169.472.4kv-cache0.264.563.642.8channel-stft57.755.547.253.5irls-path52.465.860.159.4peak-covariance25.741.016.127.6envelope-curvature44.37.349.333.6radiance-moments67.350.756.858.3kalman-update76.378.281.778.7csi-pose52.90.00.017.6vmc-update0.174.70.125.0pose-correction66.33.764.845.0spectral-reconstruction66.341.262.156.5portfolio-risk84.367.258.169.9contact-forces73.369.955.966.3density-propagation25.526.357.236.3tensor-train73.594.386.384.7