npr · DeepSeek V4 Pro run 1
34.7Library score
Pass rate82.4%
Simplicity46.3
Problems $$8.08
Library $$0.31
The library
npr · designed by DeepSeek V4 Pro (mini-SWE)2,766 source lines
ImplementerPass rateSimplicityCostScore
DeepSeek V4.1 Flash · mini-SWE85.7%41.2$2.05
GLM 5.3 Flash · mini-SWE77.5%46.1$4.58
GPT-5.6 Luna · Codex84.0%51.5$1.44
ProblemDeepSeek V4.1 FlashGLM 5.3 FlashGPT-5.6 LunaMean
cluster-consensus18.610.827.819.0logistic-path35.539.390.755.2relu-regions9.19.914.411.1table-reasoning22.245.256.941.4embedding-pooling1.256.759.039.0kv-cache52.60.20.117.7channel-stft55.047.543.648.7irls-path52.358.452.454.4peak-covariance1.128.016.715.3envelope-curvature47.045.347.946.7radiance-moments52.860.362.358.4kalman-update36.843.752.544.3csi-pose42.50.00.014.2vmc-update0.166.50.122.3pose-correction50.44.652.035.7spectral-reconstruction23.516.219.119.6portfolio-risk59.053.375.562.6contact-forces29.333.032.331.5density-propagation17.14.540.320.6tensor-train38.80.167.435.5