npr · DeepSeek V4 Pro run 3
34.3Library score
Pass rate83.7%
Simplicity45.0
Problems $$8.97
Library $$0.36
The library
npr · designed by DeepSeek V4 Pro (mini-SWE)3,067 source lines
ImplementerPass rateSimplicityCostScore
DeepSeek V4.1 Flash · mini-SWE93.9%42.8$2.62
GLM 5.3 Flash · mini-SWE81.8%45.0$4.58
GPT-5.6 Luna · Codex75.4%47.3$1.77
ProblemDeepSeek V4.1 FlashGLM 5.3 FlashGPT-5.6 LunaMean
cluster-consensus32.18.36.515.6logistic-path31.955.562.550.0relu-regions20.610.115.815.5table-reasoning87.436.361.861.8embedding-pooling44.760.470.558.5kv-cache46.547.144.946.2channel-stft45.444.241.743.7irls-path46.955.143.748.6peak-covariance23.531.622.625.9envelope-curvature35.06.239.026.7radiance-moments51.951.012.638.5kalman-update47.756.954.353.0csi-pose36.535.60.024.0vmc-update0.10.10.10.1pose-correction45.955.650.850.7spectral-reconstruction23.421.322.022.3portfolio-risk48.549.258.452.0contact-forces27.829.011.722.8density-propagation15.75.713.611.7tensor-train56.70.00.419.0